A Multi-Agent and Attention-Aware Enhanced CNN-BiLSTM Model for Human Activity Recognition for Enhanced Disability...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning
• Human Activity Recognition

Background:

• AI-driven Human Activity Recognition (HAR) is crucial for assistive technologies, aiding in fall detection, rehabilitation progress tracking, and personalized movement analysis.
• HAR applications extend to industries like surveillance, sports analytics, and medical diagnosis, highlighting its broad impact.

Purpose of the Study:

• To propose a novel, accurate, and automated strategy for human activity recognition using a hybrid deep learning (DL) and machine learning (ML) approach.
• To develop an ensemble model, Attention-CNN-BiLSTM with selective ML, for enhanced activity classification performance.

Main Methods:

• A three-stage feature ensemble strategy combining DL and ML models.
• Enhancement of Convolutional Neural Network (CNN) and Bi-directional Long Short-Term Memory (BiLSTM) with selective ML classifiers and an attention mechanism.
• Utilized publicly available datasets (UCI-HAR and WISDM) for evaluation.

Main Results:

• The proposed Attention-CNN-BiLSTM with selective ML model achieved superior predictive accuracy of 98.75% on the UCI-HAR dataset and 99.58% on the WISDM dataset.
• The ensemble approach outperformed individual models such as CNN, LSTM, CNN-BiLSTM, and Attention-CNN-BiLSTM in effectiveness, accuracy, and practicability.
• Top-performing ML and DL models were selected and combined in three stages for optimal feature extraction.

Conclusions:

• The developed comprehensive activity recognition system demonstrates significant potential for improving assistive technologies.
• The system can be integrated into advanced disability monitoring and diagnosis systems to enable predictive assistance and personalized rehabilitation strategies.