A Comparative Analysis of Hybrid Deep Learning Models for Human Activity Recognition

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning
• Human Behavior Analysis

Background:

• Human Activity Recognition (HAR) is crucial for supporting elderly care and individuals with cognitive disorders.
• Deep learning (DL) models, particularly Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), have shown high accuracy in HAR.
• Existing research highlights the potential of DL in analyzing human behavior for various applications.

Purpose of the Study:

• To investigate the effectiveness of integrating CNNs with various RNN architectures for Human Activity Recognition.
• To analyze the performance of four hybrid CNN-RNN models: CNN-LSTM, CNN-BiLSTM, CNN-GRU, and CNN-BiGRU.
• To evaluate the proposed models on the PAMAP2 dataset for daily activity recognition.

Main Methods:

• Development and analysis of four hybrid deep learning models combining CNNs with LSTMs, BiLSTMs, GRUs, and BiGRUs.
• Utilizing the PAMAP2 dataset for training and evaluating the performance of the hybrid models.
• Assessing model performance using key metrics such as F-score, accuracy, sensitivity, and specificity.

Main Results:

• The proposed hybrid CNN-RNN models achieved outstanding performance in Human Activity Recognition.
• Integration of CNNs with RNN variants (LSTMs, BiLSTMs, GRUs, BiGRUs) demonstrated significant improvements in activity recognition accuracy.
• Experimental results on the PAMAP2 dataset confirmed the effectiveness of the hybrid approaches.

Conclusions:

• Hybrid deep learning models integrating CNNs and RNNs offer a powerful approach for accurate Human Activity Recognition.
• These advanced HAR methods have significant implications for healthcare, assistive technologies, and other application domains.
• The study validates the superior performance of integrated CNN-RNN architectures over individual models for complex activity recognition tasks.