Subcarrier Selection for HAR Using CSI and CNN: Reducing Complexity and Enhancing Accuracy

Human Activity Recognition (HAR) using Channel State Information (CSI) is a promising approach for healthcare monitoring and smart environment applications. However, CSI estimation in Orthogonal Frequency Division Multiplexing (OFDM) can lead to redundancy and increased computational complexity. In this study, we propose a subcarrier selection method that leverages statistical properties, such as non-stationarity and coefficient of variation, to focus on subcarriers with the most relevant data for CSI estimation. A Convolutional Neural Network (CNN) architecture is then used to process the selected subcarriers for accurate classification. By applying our method to the eHealth CSI dataset, we significantly improve classification accuracy while reducing the computational burden. Our results show that selecting 20 to 40 subcarriers yields optimal performance, achieving a superior accuracy of 98% if compared to approaches utilizing 60 or more subcarriers. This work also presents a more efficient and scalable solution for HAR using CSI, with potential applications in real-time monitoring and low-cost systems.