Geometry-Guided Self-Supervised Learning for Label-Free 3D Indoor Radio Sensing

초록

In this paper, we propose geometry-guided selfsupervised learning (Geo-SSL), which serves as a label-free framework for mapping 3D reflection points in indoor integrated sensing and communication (ISAC) systems. To overcome the need for costly 3D ground-truth labels in supervised models and the noise sensitivity of classical geometric methods, Geo-SSL directly maps noisy path parameters, specifically angle of arrival (AoA) and time of flight (ToF), to 3D reflection coordinates. By minimizing a geometric consistency loss, the network intrinsically suppresses measurement uncertainty without explicit labels. Simulation results under modeled pathparameter uncertainty, introduced as Gaussian noise on raytracing-derived path-parameters, demonstrate that the proposed approach achieves sub-decimeter accuracy. Notably, it reduces the root-mean-square error (RMSE) by 75.2% compared to the conventional Levenberg-Marquardt (LM) baseline, providing a label-free mapping solution with performance comparable to supervised models.

키워드