LWiHS: A Lightweight Wi-Fi-Enabled Human Sensing Using Feature Fusion Strategy

Human sensing based on Wi-Fi channel state information (CSI) has attracted attention due to its nonintrusiveness and wide applications. However, due to its susceptibility to external environmental noise and interference, high-complexity deep learning (DL) algorithms are usually adopted to improve the sensing accuracy, which can hardly be deployed in edge devices with limited computational abilities. Building a high-precision, low-complexity human sensing system remains challenging. To improve the accuracy of CSI time series (TS) classification, we proposed a TS-to-image conversion method based on feature fusion, which is developed to increase the spatial structure information of the original data and expand the feature vector space. To build an efficient and lightweight feature extraction network, we designed a framework LWiHS, which is a lightweight universal feature extraction network used to extract the perceptual information containing both deep and shallow features in the images. To reduce the number of model parameters, we adopted channel pruning based on layer adaptive amplitude pruning (LAMP) scoring. LWiHS is lightweight while maintaining good feature extraction capabilities. Our proposed LWiHS outperforms other advanced algorithms in both complexity and sensing performance on four open-source datasets.