Multi-Frequency RF Sensor Human Activity Recognition

Overview

This project tackles human activity recognition (HAR) using radio-frequency sensing rather than cameras or wearables. RF-based approaches avoid the privacy concerns of cameras and aren't affected by temperature or lighting, making them a strong candidate for indoor HAR. The system fuses signals from three different sensors across 11 activities and gaits in a multimodal deep learning architecture, then quantizes the model for edge deployment.

Approach & results

• Multimodal architecture fusing three RF sensors; spectrogram-style inputs resized to 128×128 and split into train / val / test.
• Trained for 20 epochs (categorical cross-entropy, Adam) — 99% training and 89% validation accuracy.
• Float-16 post-training quantization with TensorFlow Lite cut model size ~6× (15.1 MB → 2.5 MB) and sped up execution for edge inference.