CALGA-Net: Convolution–attention with local–global aggregation for robust CSI-based indoor localization

Abstract

The performance of Wi-Fi CSI fingerprinting is often degraded by non-line-of-sight (NLOS) propagation, environmental dynamics, and device heterogeneity. To overcome these challenges, this work introduces CALGA-Net, an end-to-end coordinate regression framework designed for robust indoor localization. At its core, CALGA-Net employs a dual-path attention block that operates in parallel at the same spatial scale. Each residual unit combines large-kernel convolution with self-attention to integrate local priors and long-range dependencies. A lightweight fusion mechanism aggregates the two streams while preserving spatial consistency. The attention pathway retains directional semantics and strengthens non-local correlation modeling. A multi-scale atrous spatial pyramid pooling head aggregates context before predicting two-dimensional coordinates. The framework follows an offline training and online single-pass inference paradigm and requires no geometric ranging such as AoA or ToF or auxiliary sensors. We evaluate CALGA-Net on two public CSI localization datasets across three indoor environments: Lab, Meeting, and Hallway. CALGA-Net achieves mean localization errors of 0.088 m, 0.075 m, and 0.336 m in these environments, with standard deviations of 0.084 m, 0.122 m, and 0.406 m. These results support the effectiveness of combining large-kernel convolution with attention for CSI-based indoor localization.