Edge-Aided Radar-Based Exercise Form Classification Using Lightweight Ensemble Learning for Personalized Healthcare.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Real-time, non-contact exercise form monitoring is crucial for personalized rehabilitation.
• Existing methods like cameras and wearables face privacy and robustness challenges.
• Edge computing offers a solution for on-device processing and data privacy.

Purpose of the Study:

• To develop and evaluate RehabRadar, an edge-aided system for personalized, real-time exercise form monitoring using mmWave radar.
• To classify exercise execution as correct or identify common error types.
• To optimize the system for on-device deployment on resource-constrained hardware.

Main Methods:

• Utilized a TI IWR1642 mmWave FMCW radar for data acquisition.
• Implemented a signal processing pipeline including range-gating and FIR high-pass filtering.
• Employed a hybrid ensemble model (MobileNetV3 and CNN-LSTM) for parallel inference with late fusion.
• Optimized models for a Raspberry Pi 4B using pruning and INT8 quantization.

Main Results:

• Achieved 91.5% accuracy and 92.0% macro-F1 score in a proof-of-concept study with 10 healthy adults.
• Demonstrated an average on-device end-to-end latency of 1.5 seconds for near real-time feedback.
• The final model has a footprint of 2.1 MB, with all processing occurring on-device.
• Raw radar data is discarded post-feedback, ensuring data privacy.

Conclusions:

• RehabRadar demonstrates the feasibility of personalized, on-device exercise error detection using mmWave radar under edge constraints.
• The system balances accuracy, latency, and footprint for private, contactless home rehabilitation.
• Further validation is needed for clinical efficacy and generalizability across diverse populations and exercises.
• This work advances decentralized learning in IoT healthcare for rehabilitation applications.