Advanced IoT-Driven Human Activity Recognition and Real-Time Localization using Deep Neural Networks: A Comprehensive Approach for Intelligent and Context-Aware Systems

Over the past decade, Human Activity Recognition (HAR) has advanced significantly due to smart devices, large datasets and AI breakthroughs, enabling accurate, real-time activity prediction. This study proposes HAR-CSI-NCGNN, a novel framework using Channel State Information and Node-Level Capsule Graph Neural Networks. Data collected in indoor office settings are preprocessed using the PMCKF to reduce noise and enhance signal clarity. Feature extraction is performed using SSCET, capturing key signal dynamics for radar, biomedical and speech-related tasks. The system supports cross-environment recognition while preserving user privacy and improving generalization. Classification is handled by EPTANN, whose parameters are optimized by the Bitterling Fish Optimization Algorithm (BFOA). Implemented in Python, the framework is evaluated against CNN, ANN, and GNN-based baselines. Experimental results show that EPTANN-BFOA achieves 95–100% accuracy, significantly outperforming state-of-the-art models.