A Hybrid CNN-Transformer Framework for Precise Indoor Positioning in 6G Networks
Year: 2026
Type: conference-paper
Abstract: Precise indoor localization in dense industrial environments remains a formidable challenge due to severe NonLine-of-Sight (NLoS) propagation, which degrades traditional geometric estimation methods. This paper presents FusionLoc, a robust deep learning framework for direct 6G indoor positioning using distributed Channel State Information (CSI) from multiple Base Stations (BSs). Our approach synergizes a lightweight onedimensional Convolutional Neural Network (1D-CNN) for feature extraction from complex Channel Impulse Responses (CIRs) with a geometry-aware transformer encoder that fuses multi-modal measurements-CIR, Reference Signal Received Power (RSRP) and spatial context from BS coordinates. By explicitly modeling global dependencies between distributed BSs and employing a tail-aware training strategy, FusionLoc achieves state-of-the-art accuracy on the standard 3GPP Indoor Factory-Dense Clutter, High BS (InF-DH) benchmark. We report a test-set RMSE of 0.085 m and a 90th percentile$(P_{90})$error of 0.129 m, demonstrating decimeter-level precision and sub-meter reliability essential for safety-critical Industry 4.0 applications.
Institutions: Dhirubhai Ambani University, Tejas Networks (India)
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202360102