View-spatial-attention Augmentation on Multi-view Micro-Doppler Spectra for Hidden Human Activity Recognition

Low-frequency multi-static bio-radar technology enable the through-wall detection of human behavior in hidden buildings space from multiple perspectives. Combined with multi-view information fusion technology, it is expected to achieve robust recognition of arbitrary orientation human behavior in hidden space. However, due to the anisotropy of the multi-view detection paths and the human targets, the acquired single-view micro-Doppler spectrum (MDs) has heterogeneity in both the view channel and the feature-spectrum space, which means different significance and contribution to behavior representation and will deteriorate the multi-view fusion recognition performance. To address this issue, this study innovatively introduces a dual-layer attention mechanism into the fusion recognition of heterogeneous multi-view MDs in two dimensions. Based on the constructed recognition model of dual-layer view-spatial-attention augmented multi-view deep convolutional neural networks (MVDCNN). The recognition process will adaptively focus more on important channels or regions according to the corresponding attention, while weakening the influence of noise and clutter interference, so as to realize the dynamic weight fusion recognition of heterogeneous multi-view MDs at the feature level. The experimental results of through-wall human behavior recognition using multi-static bio-radar prove the effectiveness of this method, compared with vanilla MVDCNN, the recognition accuracy is increased by 2.09%. Moreover, the fusion recognition performance of our proposed method is more better under high signal-to-noise ratio.