Pushing the Limit of CSI-based Activity Recognition: An Enhanced Approach via Packet Reconstruction

Fine-grained and complete Channel State Information (CSI) is essential for emerging CSI-based activity recognition applications. However, many probe packets collected for CSI measurements may be lost due to co-channel interferences and other malfunctions in practice, such as link interruptions, and thus limit the further applications of these CSI-based activity recognitions. To overcome this limitation, we propose an IM proved cO mpressive Sensing bA sed mIssing packet reC overy method, named IMOSAIC, to locate the lost probe packets and to reconstruct the missing CSIs, and thus improve the accuracy and the robustness of CSI-based activity recognitions. The key idea is to trace the probe packet flow to locate the positions of lost packets, derive the CSI Matrix from CSI measurements, and use improved compressive sensing technique to reconstruct the missing CSIs. We mainly address challenges in locating the lost packets, transforming CSI measurements into CSI Matrix, and digging up CSI measurement correlations and inherent low-rank properties to reconstruct the lost packets. Furthermore, experiment results show that IMOSAIC outperforms existing interpolation methods on reconstructing the lost packets, and can achieve an average recovery accuracy of 80.21%, when 90% of packets are lost, and the reconstructed CSI datasets can improve the activity recognition accuracy obviously.