The growing elderly population makes it challenging to provide safe, effective, and respectful healthcare, especially for those living alone or in assisted facilities. Existing health monitoring systems, mainly using centralized architectures and single-sensor data, face latency, limited context awareness, and privacy issues. Although deep learning and IoT systems have advanced Human Activity Recognition (HAR), most still send raw data to the cloud, which is not ideal for elderly care. This study addresses these gaps by developing a people-focused, privacy-preserving edge monitoring framework using immersive modeling and multimodal sensing. The framework uses federated learning to let people train models together without sharing their raw personal data. It combines wearable devices, ambient sensors, and vision-based data streams with federated learning to enable decentralized, collaborative model training without exposing raw personal data. Herein, privacy is further enforced through embedded Gaussian-based visual obfuscation and a proof-of-concept Gaussian splatting module for anonymized 3D human modeling. The proposed strategy has been tested on multimodal datasets like UP-Fall and has been shown to perform well. The findings indicate the existence of distinct activity signatures across modalities. Late-fusion models also show superior fall-detection recall than unimodal baselines, even when there are issues with class imbalance and noise. Overall, the findings show that edge AI that is immersive and cares about privacy might greatly improve the accuracy, speed, and dependability of health monitoring systems for older people.
