Nuclear Norm-Based Transfer Learning for Instantaneous Multi-Person Indoor Localization

Abstract

Passive indoor localization is emerging as a transformative technology in consumer electronics, notably improving applications in smart buildings, indoor navigation, and dynamic beamforming. Our proposed CSI-ResNet transcends traditional single-target approaches by achieving a multi-target localization accuracy of 99.21% with a precision of 0.6 meters using single-timestamp CSI, surpassing existing methodologies. To mitigate model degradation from WiFi hardware phase errors and the conflation of human and locational features, we implement precise phase compensation and targeted band-stop filtering. Additionally, we have developed a pre-training methodology anchored in nuclear norms that optimizes the network for low-rank representations, significantly enhancing its transferability and ensuring consistently high performance across three transfer scenarios, with accuracy metrics reaching 86.30%, 97.03%, and 93.97% respectively. Furthermore, A robust dataset across varied settings was curated, validating our model’s effectiveness and providing extensive resources for advancing CSI-based localization predictions.