Location-Independent CSI Human Activity Recognition via Common-Differential Mode Information Decoupling and Adaptive Reconstruction

In the field of CSI based human activity recognition, due to the sensitivity of wireless channels to spatial location, CSI data exhibits significant differences in different locations. Existing work attempts to improve the location generalization ability of activity recognition by fusing multiple CSI information through multiple receivers, ignoring the shared and differential information of CSI data from different receivers, resulting in limited improvement effect. To address this, we proposed a location independent human activity recognition system CDAR based on common-differential mode information decoupling and adaptive reconstruction. Specifically, CDAR proposes a parallel common-differential mode information decoupling network that decouples the spatial similarity and spatial difference of activity features by capturing the common mode component and differential mode component of two receivers, and further incorporates a cross-deformable attention mechanism to adaptively reconstruct the decoupled features, thereby enhancing the spatial generalization of the features representations. Experimental verification was conducted based on the Widar3.0 dataset, and the average activity recognition accuracy of the CDAR system under cross location conditions reached 96.24 %, confirming the effectiveness of the proposed method.