A lightweight attention deep learning method for human-vehicle recognition based on wireless sensing technology

Abstract

Wireless sensing-based human-vehicle recognition (WiHVR) methods have become a hot spot for research due to its non-invasiveness and cost-effective advantages. However, existing WiHVR methods shows limited performance and slow execution time on human-vehicle classification task. To address this issue, a lightweight wireless sensing attention-based deep learning model (LW-WADL) is proposed, which consists of a CBAM module and several depthwise separable convolution blocks in series. LW-WADL takes raw channel state information (CSI) as input, and extracts the advanced features of CSI by jointly using depthwise separable convolution and convolutional block attention mechanism (CBAM). Experimental results show that the proposed model achieves 96.26% accuracy on the constructed CSI-based dataset, and the model size is only 5.89% of the state of the art (SOTA) model. The results demonstrate that the proposed model achieves better performance on WiHVR tasks while reducing the model size compared to SOTA model.