Self-Supervised Indoor Tracking With Radio Error Map Generation for Zero-Shot Adaptation

Abstract

Accurate indoor localization and tracking are foundational to next-generation location-based services. Although data-driven techniques offer strong modeling capabilities in complex indoor environments, they typically require labeled data for training and lack generalizability to unseen environments. In this letter, we propose a unified framework that jointly performs tracking and floor plan-aware radio error map (REM) generation for zero-shot adaptation to new environments without using real measurements or labels from those environments. This framework consists of a stochastic bidirectional recurrent neural network (SBiRNN)-based tracking model (TM) that estimates position and error statistics from unlabeled range measurements, and a vision transformer (ViT)-based REM model (REMM) that predicts spatially varying error distributions from floor plans. The TM’s estimates from seen environment first serve as pseudo-labels to train the REMM, which then generates synthetic measurements to retrain the TM for zero-shot adaptation in unseen environments. Experiments demonstrate that our framework outperforms conventional tracking methods when trained on unlabeled real measurements in a seen environment, and achieves accuracy comparable to that obtained by training on real data when trained on REMM-generated data in unseen environments.