Attention-driven
BI-LSTM for Robust Human Activity Recognition and Classification in Disaster Scenarios
Pagare Shreyas, Kumar Rakesh and Gupta S.K.
Disaster Advances; Vol. 18(8); 19-32;
doi: https://doi.org/10.25303/188da019032; (2025)
Abstract
Accurate and robust human activity recognition are essential for surveillance, healthcare
and smart environment applications. However, the unpredictability and complexity
of human motions provide significant challenges in obtaining the desired levels
of accuracy and robustness. Conventional machine learning models such as Decision
Tree, Gaussian NB and K Neighbors, have shown limited efficacy with accuracy estimates
ranging from 78.3% to 89.3%. Cutting-edge techniques like Random Forest, RBF SVC
and XGB Classifier achieve a maximum accuracy of 93.8%. We introduce a BI-LSTM model
that focuses on the most important features using bi-directional long short-term
memory networks with a special attention mechanism to address these limitations.
The present model demonstrates exceptional performance, attaining an accuracy of
99.83%, a precision of 99.46%, a recall of 99.75% and an F1 score of 99.85%, thereby
surpassing other approaches by a substantial margin. The obtained findings validate
the model's resilience and effectiveness in precisely recognizing and categorizing
human actions in different fields and situations.
