Low-Cost Wi-Fi CSI Human Presence Detection in Severe NLOS via Automated Cross-Modal Labeling

Device-free human presence detection using Wi-Fi Channel State Information (CSI) provides a passive, non-intrusive alternative to traditional indoor surveillance. However, real-world deployment faces severe challenges due to expensive specialized hardware, labor-intensive manual data annotation, and signal attenuation in Non-Line-of-Sight (NLoS) environments. This study presents a low-cost, automated Wi-Fi sensing framework using ESP32-C6 microcontrollers paired with a YOLO-based cross-modal computer vision labeling pipeline. To evaluate feature extraction and accuracy trade-offs, a comprehensive suite of algorithms was benchmarked, comprising classical machine learning baselines (Random Forest, Support Vector Machines, Naïve Bayes, and K-Nearest Neighbors) and deep learning architectures (1D-CNN, 2D-CNN, and an MLP-LSTM hybrid network). The system was evaluated under both Line-of-Sight (LoS) and challenging NLoS conditions through two 18 cm thick reinforced concrete shear walls. In LoS scenarios, Random Forest achieved an offline accuracy of 96.33% using raw, uncompressed CSI data. In severe NLoS conditions, 1D-CNN achieved a validation accuracy of 99.22%, while MLP-LSTM provided superior prediction stability for real-time edge deployment on a Raspberry Pi 5 without GPU acceleration. Furthermore, a leakage-aware temporal split analysis revealed significant performance degradation on chronologically later test data, highlighting the critical impact of temporal non-stationarity in CSI signals. This work demonstrates the feasibility of ultra-low-cost, automated through-wall sensing for indoor security and tactical monitoring applications.