STAR: A Privacy-Preserving, Energy-Efficient Edge AI Framework for Human Activity Recognition via Wi-Fi CSI in Mobile...

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Human Activity Recognition (HAR) using Wi-Fi Channel State Information (CSI) is a privacy-preserving sensing method for smart homes and IoT.
• Existing CSI-based HAR methods face challenges with computational efficiency, latency, and deployment on resource-constrained devices.

Purpose of the Study:

• To develop STAR, an edge AI-optimized framework for real-time, energy-efficient HAR on low-power embedded devices.
• To address the limitations of existing CSI-based HAR approaches.

Main Methods:

• STAR integrates a lightweight three-layer Gated Recurrent Unit (GRU) for temporal modeling, reducing parameters by 33% compared to LSTM.
• A multi-stage pre-processing pipeline (median filtering, Butterworth low-pass filtering, Empirical Mode Decomposition) is used for signal denoising and feature extraction.
• The system is implemented on a Rockchip RV1126 processor with an NPU and an ESP32-S3 CSI acquisition module.

Main Results:

• Achieved 93.52% mean accuracy for seven activity classes and 99.11% for human-presence detection with a 97.6k-parameter model.
• INT8-quantized inference reached 33 MHz throughput with 8% CPU utilization, a six-fold speedup over CPU-based execution.
• Demonstrated sub-second response latency and low power consumption.

Conclusions:

• STAR provides a practical, scalable solution for real-time, privacy-preserving HAR in mobile and pervasive computing.
• The framework enables efficient on-device HAR, overcoming previous computational and latency barriers.