Device-Free Occupancy Detection via Wi-Fi CSI and Deep Residual CNN
Seonghyeon Park, Jaehan Joo, Suk Chan Kim · 2025
This paper explores a device-free indoor occupancy detection approach using raw Wi-Fi Channel State Information (CSI). CSI data, including both real and imaginary components per subcarrier, were collected over three days in a$5.75 \mathrm{m} \times 3.7$m room under empty and occupied (single-person) conditions, with 100 CSI packets per second transmitted and received for two minutes per session. A deep 1D Convolutional Neural Network (1D-CNN) with residual blocks was trained to classify the binary occupancy status. Two temporal dataset scenarios were constructed by rearranging the order of daily CSI logs to evaluate the robustness of the model against temporal variation. The proposed system utilizes low-cost ESP32 devices and demonstrates high accuracy without requiring explicit preprocessing or feature extraction, highlighting its potential for practical and scalable indoor sensing.