Description of Invention
Fingerprint-based indoor localization addresses challenges in GPS-restricted environments with minimal hardware requirements. Despite their popularity, these systems often suffer from signal fluctuations caused by shadowing, fading, and multipath effects, which undermine accuracy. While 1D and 2D CNNs can extract spatial features, they fail to capture temporal variations, thereby limiting their online localization accuracy. Although 3D CNNs is capable of extracting spatio-temporal information, their associated computationally burden hinder real-time deployment. This study proposes novel knowledge distillation (KD)-based 3D CNNs for indoor positioning , where a high-capacity 3D CNN teacher model guides lightweight student models by transferring intermediate feature representations. This method significantly enhances the student’s positioning accuracy while reducing complexity, achieving performance close to the teacher model with 65% of lower computational demands.