Gaussian fitting based human activity recognition using Wi-Fi signals

With the popularity of commercial Wi-Fi devices, channel state information (CSI) based human activity recognition shows great potential and has made great progress. However, previous researchers always tried to remove the noise signals as much as possible without considering the distribution characteristics. Different from the previous methods, we observed the phenomenon that the signal distribution is different when the action exists and does not exist, so we propose GFBR. GFBR takes noise distribution as the entry point, proposes a novel human activity modelling method, and designs a dual-threshold segmentation algorithm based on the modelling method. Then, we extract features from amplitude and linearly corrected phase to describe different activities. Finally, a support vector machine (SVM) is used to recognise five different activities. The average recognition accuracy of GFBR in the three different environments is 94.8%, 96.2%, and 95.7%, respectively, which proves its good robustness.