Non-Contact CSI-Based Behavioral Sensing Using Latent Representation and Ensemble Transformer

Abstract

Non-contact human activity recognition (HAR) is attracting increasing attention since it shows promise for various Human-computer Interaction (HCI) applications. However, most of HAR solutions are vulnerable to channel dynamics in indoor settings. To address this issue, a non-contact HAR scheme based on ensemble Transformer is developed in this letter. The main idea is to measure and analyze channel state information (CSI) of wireless signal reflected off a user to identify potential activities. Specifically, we design a multi-view CSI fusion method (MVCF) to calculate key feature by aggregating all CSI measurements. On this basis, we propose an ensemble learning model by combing classification blocks with lightweight voting strategy. Each block is comprised of a Transformer encoder and Softmax regression, and aims to learn high-level abstract representation from the key feature for capturing reliable mapping between the CSI and daily activities. Additionally, the voting strategy is designed and used to determine the activity label for each unknown instance by aggregating classification results from all blocks. Experimental results show that our scheme can achieve superior performance with an average recognition accuracy of 98.3%.