Towards ubiquitous fingerprint-based localization with limited human effort

Recent fingerprint-based localization in a deep-learning framework has promising potential for providing location-based services with 6G’s inherent intelligence. Classical deep neural networks are designed to extract features that can effectively improve accuracy in specific scenarios. However, the hurdles of radio annotation, fingerprint degradation, and feature dependence severely limit its universal applicability with unpredictable environmental dynamics. To address these issues, we propose a novel deep-learning-based localization framework to achieve ubiquitous location estimation with minimal human effort. The latent view-invariant factor underlying channel state information (CSI) data is refined via amplitude and phase sampling to establish a dual-view contrastive pretraining model, DVCLoc, that learns generic representations directly transferable to new environments or scenarios. Then, limited CSI fingerprints are used to train the location predictor for robust localization. Extensive real-world experiments demonstrate that DVCLoc achieves state-of-the-art localization performance across various complex scenarios, advancing DNN-based localization from specificity to generality.