A Robust Signal Processing Framework for CSI-Based Bangla Sign Language Recognition via Doppler-Based Subcarrier Selection, Complex ICA, and Energy-Based Fusion

WiFi Channel State Information (CSI) has emerged as a promising medium for device-free human activity recognition due to its ubiquity and privacy-preserving nature. However, existing CSI-based gesture recognition frameworks have primarily focused on coarse activities or dominant sign languages like ASL, leaving regional languages such as Bangladeshi Sign Language (BaSL) largely unexplored. In this work, we propose the first WiFi CSI-based framework for BaSL word-level gesture recognition. Our system introduces a novel preprocessing pipeline, SCANEM (Subcarrier Selection, Complex independent component Analysis, and Energy-guided Motion Fusion), which integrates Doppler-based subcarrier filtering, Complex independent component analysis, and energy-guided motion fusion to extract gesture-relevant patterns while mitigating gain-phase distortions. Experimental evaluation on an WiBaSL dataset demonstrates that our two-block CNN model—powered by the proposed preprocessing pipeline—achieves 89.62% accuracy and an F1-score of 0.88, while maintaining low model complexity. Our preprocessing framework also addresses challenges in multi-user scenarios and provides a deployable path toward inclusive and language-diverse gesture recognition systems.