Federated Transfer Learning on the Edge: A Vision-Based Motor Rehabilitation System

• Published in: International Conference on Distributed Computing in Sensor Systems and workshops (Online) pp. 523 - 530
• Main Authors: Esposito, Alfonso, Moghbelan, Yasamin, Zyrianoff, Ivan, Bononi, Luciano, Di Felice, Marco
• Format: Conference Proceeding
• Language: English
• Published: IEEE 09.06.2025
• Subjects: Accuracy; Edge intelligence; Federated learning; Human Activity Recognition; Internet of Things; Monitoring; Motors; Privacy; Smart systems; Training; Training data; Transfer learning
• ISSN: 2325-2944
• DOI: 10.1109/DCOSS-IoT65416.2025.00086

Integrating IT solutions to support physical rehabilitation exercises can significantly improve health outcomes. Among these, Computer Vision (CV)-based solutions stand out as promising tools for automating exercise monitoring while remaining non-intrusive. However, CV techniques that rely on Deep Learning (DL) algorithms require extensive datasets and dedicated infrastructure. Further, centralized training data expose sensitive data, while local-only training risks poor generalization and low accuracy when encountering new subjects. To address these limitations, we propose FLORE (Federated Transfer Learning for Optimized Recognition on the Edge), a novel framework designed to support edge-friendly, privacypreserving motor routine classification during rehabilitation sessions. The architecture leverages Transfer Learning (TL) for efficient feature extraction from video frames and Federated Learning (FL) for collaborative model training across distributed devices, ensuring patient privacy. The DL models in FLORE are both lightweight and quantized, as demonstrated by our experiments, making them well-suited for resource-constrained edge environments as demonstrated by our experiments in which we successfully deployed the framework in five different edge devices. Our results demonstrate that FL solutions achieve the best generalization across all experiments while we drastically reduce the model size (from 10MB to <4 \text{MB} ), still maintaining a similar accuracy.