In-Home Human Activity Recognition via Kinematics-Focused Multimodal Sensor Fusion and Spatio-Temporal Neural Architecture

This letter presents a robust human activity recognition (HAR) framework for smart home environments, using multimodal sensor fusion and advanced spatio-temporal deep learning modeling. The system integrates data from a wrist-worn actigraphy device, a real-time location system, and a vision-based activity sensor. A hybrid deep learning model combining convolutional neural networks, bidirectional long short-term memory, and neural structured learning, enhanced by an attention mechanism, enables accurate interpretation of complex human activity patterns. The results demonstrate the efficacy of multimodal sensor fusion and spatio-temporal neural architectures that enhance activity recognition accuracy. The model achieved 95.19% accuracy for routine daily activities, outperforming conventional nontemporal models due to its ability to handle the complexity inherent in spatio-temporal data effectively.