Optimizing Deep LSTM Model through Hyperparameter Tuning for Sensor-Based Human Activity Recognition in Smart Home

Abstract

Human Activity Recognition (HAR) holds significant potential in healthcare, smart homes, sports, and security, mainly benefiting the well-being of elderly individuals and dependents. This research introduces an innovative deep learning-based approach to HAR, using wearable sensors in smart home environments. In this paper, we conduct a comprehensive review of the state of the art, offering insights into existing methods, classification techniques, their performances, hyperparameter tuning strategies, findings, limitations, and future directions. We propose an LSTM-based deep model enriched with batch normalization and perform a hyperparameter tuning using Bayesian Optimization; then, we evaluate the model on the PAMAP2 public dataset. The model outperforms previous studies, achieving remarkable performance metrics, including accuracy at 97.71%, F1 score, precision, and recall, approaching 96.66%, 96.85%, and 96.55%, respectively. We plan to assess the model's generalization capabilities for future work by training it on diverse datasets such as Opportunity and WISDM. Furthermore, we aim to enhance the model by exploring hybrid deep model architectures and alternative hyperparameter tuning approaches. These efforts maximize the model's efficiency and adaptability in real-world scenarios.