Hierarchical Contrastive Learning for CSI-Based Fingerprint Localization

Abstract

Fingerprint localization based on Channel State Information (CSI) plays a crucial role in indoor location-based services. Due to the natural compatibility between offline training and online localization of CSI-based fingerprint localization and deep learning, recent studies have shown that introducing the latest deep learning techniques can provide higher localization accuracy. Most current research efforts in localization have focused on leveraging deep learning advancements to enhance performance. However, these approaches typically rely on complex techniques and large model sizes, prioritizing model-driven methods over practicality and real-world deployment capabilities. In this paper, we aim to improve the localization performance of simple, general-purpose models (e.g., $\textsf{ResNet}$) through data-driven training paradigms, which align with the value proposition of real-world applications. Specifically, by constructing positive examples with different signal-to-noise ratios (SNRs) for contrastive learning, $\textsf{ResNet}$ can learn SNR-robust representations. Furthermore, we focus on antenna instances (physical components of CSI) at a smaller granularity to learn scale-invariant representations through hierarchical loss. In the final location regression fine-tuning process, only a pooling layer and a fully connected layer need to be added to perform position mapping. Experiments on real-world indoor and urban canyon datasets demonstrate that our method achieves positioning accuracies of 0.16 m and 0.54 m, respectively, significantly outperforms state-of-the-art baseline models.