Intelligent Wireless Walls for Contactless In-Home Monitoring

The study addresses the challenge of accurate, privacy-preserving, contactless human activity monitoring in real-world, non-line-of-sight (NLoS) environments. Conventional microwave sensing suffers from limited range (weak target reflections versus strong LOS components), environmental interference, and the need for controlled, ideal settings. The authors propose Intelligent Wireless Walls (IWW), which combine reconfigurable intelligent surfaces (RIS) capable of beam steering/beamforming with machine learning-based classification to enhance reflected signal strength from human targets and suppress interference. The purpose is to extend sensing coverage and improve activity recognition accuracy in complex scenarios such as around corners (corridor junctions) and across building floors, supporting applications like remote healthcare, intrusion detection, and independent living.

Prior work on activity recognition includes ambient sensors, cameras, and wearables, each with privacy or usability drawbacks. Contactless microwave sensing (e.g., using WiFi, 5G CSI, or radar Doppler) can infer motion from reflected signal variations but struggles in NLoS due to weak reflections and interference. RIS technology—electromagnetic metasurfaces with tunable subwavelength unit cells—offers low-complexity, passive, scalable EM wavefront manipulation compared with phased arrays or relays. Recent AI-enabled RIS systems have demonstrated gesture recognition, learned sensing, and metasurface imaging using neural networks. Related NLoS imaging/tracking methods span acoustics (requiring multiple measurement positions), long-wave IR (needs IR cameras), standard cameras with lasers (privacy concerns), passive optical coherence (limited in low light), and radar using reflections/diffractions with scene geometry. A RIS-aided monostatic radar study showed SNR/SCR gains with increased RIS size. Building on these, the present work experimentally investigates an ambient-sensing-based, bistatic activity monitoring setup using RIS to create a virtual LOS in NLoS scenarios.

Experimental concept: Deploy an RIS to manipulate ambient microwave propagation by beam steering/forming, creating a virtual LOS path between a transmitter (Tx) and receiver (Rx) separated by blockages. Evaluate the effect on contactless activity monitoring using machine learning classifiers. Scenarios: (1) Corridor junction: Tx and Rx in separate corridor sections (around a corner). (2) Multi-floor: Tx and Rx on different floors of a building. Activities and subjects: Two participants (one male, one female) perform three activities: sitting, standing, and walking. An additional 'empty' class (no person) is also considered in analysis/figures, yielding seven classes: Empty, SittingS1, SittingS2, StandingS1, StandingS2, WalkingS1, WalkingS2. RIS hardware and operation: RIS with high-resolution azimuthal beam steering, previously shown to enhance indoor coverage in NLoS communications. The metasurface comprises columns of sub-wavelength unit cells, each integrating three PIN diodes to achieve near 3-bit phase resolution. Control is via WiFi for remote configuration. Unit cell design (operating at 3.75 GHz): five copper patches on grounded F4BM-2 substrate (εr = 2.65, tanδ = 0.001), interconnected by three PIN diodes and a capacitor. Forward-biased PIN diodes behave as small series resistors; reverse-biased act as series capacitors. Eight bias states (000–111) from the three diodes; reverse bias at 0 V, forward bias ~0.85 V with 3 mA. The capacitor operates near its self-resonance to provide an RF short while isolating DC bias. Adjacent unit cells are connected vertically to simplify biasing, prioritizing azimuthal control. Machine learning and evaluation: Three classifiers—Random Forest (RF), Extra Trees (ET), and Multilayer Perceptron (MLP)—are trained to classify activities. Two evaluation protocols are used: (i) train–test split and (ii) repeated stratified k-fold cross-validation. Performance is reported for each subject (S1, S2) and combined (S1+S2), with and without RIS. Confusion matrices are analyzed to understand class-wise performance with RIS-off versus RIS-on. Procedure: For each scenario, data are collected under RIS-off (conventional NLoS microwave sensing) and RIS-on (RIS optimized to form a virtual link between Tx and Rx). Classifiers are trained and evaluated per protocol; accuracies and confusion matrices are reported to quantify performance gains due to RIS.

Overall, enabling the RIS substantially improves activity recognition accuracy in both NLoS scenarios, with reported maximum detection gains of 28% (multi-floor) and 25% (corridor junction) over conventional microwave sensing. Corridor junction (train–test): - RF: S1 93.75%→100%; S2 93.75%→100%; Combined 75.00%→100.00%. - ET: S1 93.75%→100%; S2 95.47%→100%; Combined 71.87%→96.87%. - MLP: S1 93.75%→100%; S2 75.00%→100%; Combined 75.00%→87.50%. Corridor junction (repeated stratified k-fold): - RF: S1 85.66%→100%; S2 91.10%→100%; Combined 81.80%→89.58%. - ET: S1 90.81%→100%; S2 94.27%→100%; Combined 83.53%→91.47%. - MLP: S1 80.23%→99.58%; S2 86.01%→97.38%; Combined 75.04%→89.95%. Confusion analysis (corridor): Without RIS many walking activities are misclassified; with RIS-on, most classes achieve 100% accuracy; the worst case is SittingS1 with only 20% misclassification. Multi-floor (train–test): - RF: S1 63.00%→82.50%; S2 70.00%→79.50%; Combined 51.42%→72.14% (+20.72%). - ET: S1 70.00%→92.50%; S2 80.00%→85.00%; Combined 54.28%→81.42% (+27.14%). - MLP: S1 62.50%→57.50%; S2 77.50%→91.66%; Combined 52.85%→54.28%. Multi-floor (repeated stratified k-fold): - RF: S1 57.79%→78.51%; S2 73.74%→84.86%; Combined 51.38%→69.25% (+17.87%). - ET: S1 64.01%→84.28%; S2 75.39%→86.52%; Combined 55.67%→85.71%. - MLP: S1 62.63%→65.54%; S2 73.89%→79.92%; Combined 52.25%→49.59%. Confusion analysis (multi-floor): Without RIS, standing and walking for S2 are often misclassified; with RIS-on, accuracy improves across classes, with the most confusions for SittingS2 (about 26% incorrect reported). Figure-level comparisons indicate consistently positive gains with RIS, with the largest improvement observed for ET in the multi-floor scenario (reported >25%).

The results demonstrate that integrating an RIS into a contactless microwave sensing system can create a virtual LOS in otherwise NLoS settings, enabling stronger and more directed reflections from human targets and suppressing environmental interference. This beamforming/beam-steering capability directly translates into higher-quality sensing features and markedly improved machine learning classification performance. In both corridor junction and multi-floor scenarios, RIS-on conditions improve accuracies for individual subjects and combined datasets, often reaching near-perfect classification in the corridor scenario. These findings validate the IWW concept as a practical approach to extend the operational range and reliability of in-home activity monitoring without compromising privacy and without requiring subjects to wear devices. The improvements are particularly notable in the more challenging multi-floor setup, underscoring the RIS’s value in complex propagation environments.

This work introduces Intelligent Wireless Walls (IWW), combining reconfigurable intelligent surfaces with machine learning to enable high-precision, privacy-preserving, contactless human activity monitoring in challenging NLoS environments. Experiments in corridor-junction and multi-floor scenarios show substantial accuracy gains with RIS, achieving up to 28% improvement in multi-floor and 25% in corridor junction over conventional sensing, and near-perfect recognition in some corridor cases. The study highlights the feasibility of RIS to extend sensing coverage and robustness for applications like independent living and remote healthcare. Future research could scale up participant numbers and activity types, investigate diverse indoor layouts and materials, develop online/adaptive beam control and classifier training, explore multi-RIS deployments and elevation control, and assess generalization across hardware and frequency bands.

The experiments involve only two participants and three activities, limiting generalizability across populations and behavior diversity. Performance is evaluated in two specific indoor NLoS scenarios; results may vary in different environments (room geometries, materials, clutter). Gains depend on effective RIS placement and optimization; suboptimal configurations could reduce performance. Classifier performance varies by algorithm and dataset, with occasional degradations (e.g., MLP in some multi-floor combined evaluations), indicating sensitivity to model choice and data characteristics.