Neural 5G Indoor Localization with IMU Supervision

TL;DR

This work tackles indoor localization using 5G CSI by removing the need for dense ground-truth position labels. It introduces IMU-based pseudo-labels derived from double integration and a trajectory-fitting framework that uses a few control points to refine these labels, subsequently training a neural network to map CSI to positions. A forward-backward integration scheme and iterative refinement are employed to reduce drift and improve accuracy, achieving decimeter-level performance on both simulated and real-world data. The approach significantly lowers deployment effort while delivering competitive accuracy, suggesting practical viability for privacy-preserving, label-efficient localization in indoor environments.

Abstract

Radio signals are well suited for user localization because they are ubiquitous, can operate in the dark and maintain privacy. Many prior works learn mappings between channel state information (CSI) and position fully-supervised. However, that approach relies on position labels which are very expensive to acquire. In this work, this requirement is relaxed by using pseudo-labels during deployment, which are calculated from an inertial measurement unit (IMU). We propose practical algorithms for IMU double integration and training of the localization system. We show decimeter-level accuracy on simulated and challenging real data of 5G measurements. Our IMU-supervised method performs similarly to fully-supervised, but requires much less effort to deploy.

Figures (11)

• Figure 1: Two stages of our method. Note that during operation the position prediction does not depend on past or future steps, only observed CSI is used.
• Figure 2: Training pipeline.
• Figure 3: LoS alignment procedure. The LoS peak is detected and CIR is rolled to align the peak to a specific bin. The same shift applies to all TRPs and antennas.
• Figure 4: Forward-backward algorithm. The scheme depicts the trajectory fitting based on misalignment of the positions obtained with forward and backward integration passes. Respectively, each position of the trajectory is represented with two points. We optimize the correction of the acceleration in each position pulling these two points to each other (shown with red arrow).
• Figure 5: Simulated trajectories. Gray circles show control points, which act as starting points for integration. Red and blue trajectories are obtained with forward and backward integration. They quickly diverge from the ground truth due to the error accumulation. Green trajectory is the result of the forward-backward algorithm, it corrects the error pushing the trajectory closer to the ground truth.