SA-Loc: Scenario Adaptive Localization in Highly Dynamic Environment using Adversarial Regressive Domain Adaptation

Abstract

Localization based on signals from cellular base stations (BSs) or WiFi access points is important particularly in urban environments where street canyon effects impair GPS reception. A popular implementation is based on fingerprinting, where associations between channel characteristics and location of a moving terminal (MT) are measured in a training phase, and later (in the testing of operating phase) used to localize an MT measuring a particular channel. This paper considers the combination of labeled and unlabeled channel state information (CSI) for such finger-printing based localization, in particular to account for the fact that moving objects like cars can change the CSI between training and testing. To enhance the robustness of the localization against these and other “unseen” scenarios, we propose a solution approach based on a deep domain adaptation (DA) framework. The proposed method contains an unsupervised domain adaptation regression and an adversarial training module. The first component aims to learn discriminative fingerprinting representations jointly with data collected from seen and unseen scenes. The second one alleviates the distribution divergence of fingerprints collected from seen and unseen scenes, which eventually enables effective knowledge transfer. Finally, we evaluate the performance of the proposed method in a challenging setup, the recently released dynamic scenario in the DeepMIMO dataset. The results demonstrate that the proposed method can provide robust and accurate location estimation over variable numbers of the time-varying scenarios, with superior performance over other compared methods.