Research Foundation
WiSense is grounded in peer-reviewed WiFi sensing research. Our pipeline follows the methodology established by leading CSI-based detection systems - Widar3.0, MM-Fi, and RF-Pose - adapted for real time demonstration without dedicated hardware.
1
Signal Modeling
We model CSI amplitude across N subcarriers as a superposition of the base multipath signal and a Gaussian disturbance term caused by object presence. This follows the first-order RF propagation model used in Widar3.0.
2
Preproccessing Chain
Raw signal → median filter (spike removal) → mean subtraction (DC offset) → Butterworth low-pass filter (noise) → outlier clipping → minmax normalization. Identical to the preprocessing pipeline in WiGest and Widar.
3
2D Spatial Reconstruction
CSI disturbance is modelled as a 2D Gaussian blob in a 64×64 spatial grid, where blob position encodes object location and blob amplitude encodes object size. This is a simplified version of MUSIC-based spatial reconstruction.
4
Fixed Scale Edge Detection
Critical design decision: we use a fixed amplitude scale [0, 1.2] instead of NORM_MINMAX before Canny. This prevents empty-room noise from being stretched to full contrast — ensuring edge count is physically meaningful.
5
Dual-Classifier Ensemble
Rule-based detector (edge count threshold) provides an interpretable baseline. SVM classifier provides probabilistic output. Agreement between both increases confidence. Disagreement triggers conservative 'uncertain' output.
References
Proposes a large-scale benchmark dataset for WiFi CSI-based gesture recognition using Intel 5300 NICs. Introduces body-coordinate velocity profiles (BVP) as a domain-independent feature. Our feature extraction pipeline draws directly from the amplitude+phase features validated in this work.
Introduces a multi-modal dataset combining WiFi CSI with synchronized RGB-D video and LiDAR point clouds. Enables cross-modal validation — the same approach we use to validate our simulated CSI pipeline against real-world sensing patterns.
Demonstrates human pose estimation through walls using 2.4 GHz WiFi. Establishes that RF signal distortion encodes fine-grained spatial information sufficient for skeletal tracking. Our edge detection approach targets the precursor problem — object presence detection — as a stepping stone toward this goal.
First work to use commodity WiFi RSSI for in-air gesture recognition without dedicated hardware. Validates that standard 802.11 infrastructure is sufficient for motion sensing — the key premise behind our no-hardware approach.