Wireless Perceptual Space Modeling Method for Cross-Domain Human Activity Recognition

Existing Wi-Fi perceptual recognition using Doppler Frequency Shift (DFS) can portray human activity and behavioral features, but the method is affected by the user’s movement direction, position, and other factors, resulting in large differences in the spectral modes generated by the same action, which restricts the performance and pervasiveness of the cross-domain recognition system. Based on this, the wireless sensing Spatial Frequency Field (SFF) modeling method is proposed to extract cross-domain human activity features through DFS. The key of this method is to model the wireless sensing space according to the DFS frequency, speculate the human body orientation information and correct the DFS power, calculate the power magnitude of the spatial frequency points to generate the SFF, where the spatial power field contains the multilink frequency distributions due to human body activities and contains the human body orientation information, which unifies the DFS frequency distributions of the same human body activities under the fixed orientation, and finally, design the CNN-RNN deep neural network to classify the SFF. Under the test of multiple datasets and classification tasks, the method proposed in this article has the advantages of being lightweight, fast, accurate, and universal, in which the average accuracy of intra-domain gesture recognition reaches 95.3%, the average accuracy of cross-domain gesture recognition reaches 82.8% to 91.2%, and the accuracy of gait recognition reaches 94.6%.