A Deep Learning-Enhanced CSI-Based System for High-Precision Indoor Positioning

Accurate indoor localization is critical for emerging Internet of Things and smart environment applications. Channel State Information (CSI), with its fine spatial resolution and high sensitivity to environmental changes, offers strong potential in Wi-Fi-based indoor localization. However, practical deployment is hindered by multipath interference and device heterogeneity, which degrade positioning accuracy and system robustness. To address these challenges, this paper proposes a deep learning-enhanced CSI acquisition and processing framework, termed MSRANet. The framework incorporates parallel multi-scale temporal convolutions to capture hierarchical and fine-grained features, deep residual blocks for enhanced feature representation, and a dual attention mechanism to refine frequency and spatial features. A lightweight Weighted Moving Average module is also introduced to smooth prediction outputs and improve stability. The proposed system is implemented on low-cost embedded hardware and evaluated in multiple real-world indoor environments. Experimental results show a mean absolute error of around 0.7 meters, demonstrating improved positioning accuracy and robustness compared to existing methods, while maintaining computational efficiency suitable for real-time deployment.