Human Activity Recognition with Hybrid Deep Learning: A Case Study Using CNN-LSTM and RFC-LSTM on the Human3.6M Dataset
Meng-Hing Aun, Pei-Ching Yang · 2025
Human Activity Recognition (HAR) is essential for numerous real-world applications, including healthcare monitoring, intelligent surveillance, and innovative environments. This study compares two hybrid deep learning models—Convolutional Neural Network combined with Long Short-Term Memory (CNN-LSTM) and Random Forest Classifier combined with LSTM (RFC-LSTM)—using the Human3.6M dataset. The objective is to assess and compare their performance in terms of classification accuracy, robustness, and real-time monitoring capability. Experimental results reveal that CNN-LSTM outperforms RFC-LSTM in recognizing complex motion patterns, while RFC-LSTM demonstrates greater sensitivity to activity similarity, resulting in higher misclassification in overlapping actions. Both models, however, show limitations in consistent real-time performance due to data variability, overlapping behaviors, and architectural constraints. The study highlights the importance of incorporating attention mechanisms, multi-modal data fusion, and domain adaptation to enhance recognition stability. This work contributes to the advancement of HAR systems by analyzing the strengths and limitations of each hybrid model and offering insights for future improvements in spatial-temporal feature extraction and adaptive learning design.