Low-Cost Wi-Fi CSI-Based Human Activity Recognition Using ESP32 and a Gated Recurrent Unit Classifier

Abstract: Human Activity Recognition (HAR) has traditionally relied on cameras or wearable sensors, which raise privacy and user-compliance concerns. This paper presents a device-free HAR system based on Wi-Fi Channel State Information (CSI) captured with a single low-cost ESP32 microcontroller, addressing the accessibility gap left by prior work that depends on expensive, multi-antenna, research-grade Wi-Fi hardware. Raw CSI logs were collected from four subjects performing ten activities, parsed, and audited for recording-duration quality; 13 of 150 recordings were excluded as truncated. The remaining data were converted to CSI magnitude, segmented into overlapping one-second sliding windows, and normalized using training-set statistics. A two-layer Gated Recurrent Unit (GRU) classifier with 54,346 parameters was trained with class-weighted cross-entropy loss and early stopping. On a held-out, recording-level test split of 607 windows, the model achieved 64.74% accuracy and 65.38% macro F1-score, and macro F1 rose from 59.19% to 65.38% after the data-quality audit (measured on differently sized test splits). High-motion activities were recognized well (F1 > 0.75), whereas low-motion activities were less separable, which is consistent with the limited resolution of single-antenna CSI. A controlled live test that held the Wi-Fi hotspot constant and changed only the room showed that the learned CSI signatures are room-specific, confirming the cross-environment generalization limitation reported in the literature. The work provides an accessible, reproducible baseline for low-cost CSI-based HAR and an empirical characterization of its present limitations.

Keywords: Wi-Fi Sensing; Channel State Information (CSI); Human Activity Recognition (HAR); ESP32; Gated Recurrent Unit (GRU); Deep Learning; Device-Free Sensing