Real-time, non-contact monitoring of exercise form is important for personalized rehabilitation, yet cameras and wearables raise privacy and robustness concerns. We present RehabRadar, an edge-aided system using a TI IWR1642 mmWave FMCW radar that trains personalized models to classify execution as correct or one of two common error types. The signal pipeline isolates the subject via range-gating and FIR high-pass filtering, then extracts complementary micro-Doppler spectrograms and range-Doppler sequences. A hybrid ensemble-MobileNetV3 for micro-Doppler images and a CNN-LSTM for range-Doppler sequences-performs parallel inference with late fusion. The models are optimized for a Raspberry Pi 4B via magnitude-based pruning and post-training INT8 quantization. We evaluated the system in a proof-of-concept study with 10 healthy adults performing eight clinician-curated exercises; per-participant two-fold cross-validation yields 91.5% accuracy and 92.0% macro-F1, outperforming single-branch baselines. On-device end-to-end latency averages 1.5 s, enabling near real-time, rep-level feedback while preserving data locally. The final model footprint is 2.1 MB, and all processing occurs on-device; raw radar is discarded after feedback, and only minimal encrypted metrics are transmitted externally. This work demonstrates the feasibility of personalized on-device error detection under edge constraints. Establishing clinical efficacy and generalizability beyond the studied cohort, broader exercise taxonomies, and patient populations is outside the present scope and motivates planned multi-site clinical validation. These results indicate that the mmWave radar, combined with compact neural models, can balance accuracy, latency, and footprint for private, contactless home rehabilitation analytics, and advance decentralized learning in IoT healthcare.
Gomez et al. (Thu,) studied this question.