Abstract

Abstract
Device-free Human Activity Recognition (HAR) presents a significant challenge, offering a privacy-preserving alternative to vision-based systems. This work proposes a novel methodology that leverages the rich motion dynamics captured in Doppler traces derived from Channel State Information (CSI). We introduce a hybrid deep learning architecture, Inception-iLSTM, specifically engineered to process these traces. The Inception module excels at extracting salient, multi-scale local features from the Doppler data, while the Bidirectional Long Short-Term Memory (BiLSTM) network subsequently models the long-range temporal dependencies inherent in complex human activities. To further enhance classification performance, a Support Vector Machine (SVM) with a non-linear kernel is integrated as a post-processing stage. This step refines the decision boundaries learned by the deep neural network, significantly improving generalization. The proposed methodology achieves outstanding accuracy rates approaching 99% in identifying distinct human movements. These results are validated through comprehensive performance metrics, including confusion matrices, confirming the robustness and high efficacy of this hybrid approach for CSI-based HAR.