Wi-Sense: A Lightweight Deep Learning Framework for Device-Free Intrusion Detection and Occupancy Estimation Using Wi-Fi CSI with ESP32 Edge Deployment

The increasing demand for privacy-preserving smart surveillance systems has driven research toward device- free sensing approaches that leverage existing wireless infras- tructure. Traditional camera-based intrusion detection systems raise privacy concerns and require controlled lighting conditions, while Passive Infrared (PIR) sensors suffer from limited spatial sensitivity and high false alarm rates. This paper presents Wi- Sense, a lightweight deep learning framework for device-free physical intrusion detection and multi-class occupancy estimation using Wi-Fi Channel State Information (CSI). The proposed system captures fine-grained CSI amplitude variations using an ESP32-based CSI receiver and processes the data through a struc- tured signal preprocessing pipeline including outlier removal, smoothing, normalization, and sliding window segmentation. A compact one-dimensional Convolutional Neural Network (1D- CNN) is designed to extract spatial correlations across subcarri- ers and classify occupancy levels ranging from 0 to 5 persons. Bi- nary intrusion detection is derived from occupancy classification results. Experimental evaluation in a realistic furnished indoor environment demonstrates 96.8% intrusion detection accuracy and 92.4% occupancy estimation accuracy. Furthermore, the architecture is optimized for TinyML deployment, enabling standalone inference and alert transmission directly from the ESP32 microcontroller without reliance on external computation. The results validate the feasibility of scalable, low-cost, and privacy-aware smart building security using commodity Wi-Fi infrastructure.