[Paper Review] Position Tracking using Likelihood Modeling of Channel Features with Gaussian Processes

Sebastian Kram, Christopher Kraus+4 more|arXiv (Cornell University)|Mar 24, 2022

Indoor and Outdoor Localization Technologies4 citations

TL;DR

This paper proposes a Gaussian process regression (GPR)-based fingerprinting framework that models channel feature likelihoods for robust position tracking in sparse indoor environments. By compressing channel measurements via an autoencoder and regressing distance-dependent GPs on features, it enables statistical observation modeling for particle filter fusion, achieving 0.52 m MAE on sparse data—significantly outperforming CNN-based fingerprinting (1.3 m MAE) in low-LOS, multipath-dominant industrial settings.

ABSTRACT

Recent localization frameworks exploit spatial information of complex channel measurements (CMs) to estimate accurate positions even in multipath propagation scenarios. State-of-the art CM fingerprinting(FP)-based methods employ convolutional neural networks (CNN) to extract the spatial information. However, they need spatially dense data sets (associated with high acquisition and maintenance efforts) to work well -- which is rarely the case in practical applications. If such data is not available (or its quality is low), we cannot compensate the performance degradation of CNN-based FP as they do not provide statistical position estimates, which prevents a fusion with other sources of information on the observation level. We propose a novel localization framework that adapts well to sparse datasets that only contain CMs of specific areas within the environment with strong multipath propagation. Our framework compresses CMs into informative features to unravel spatial information. It then regresses Gaussian processes (GPs) for each of them, which imply statistical observation models based on distance-dependent covariance kernels. Our framework combines the trained GPs with line-of-sight ranges and a dynamics model in a particle filter. Our measurements show that our approach outperforms state-of-the-art CNN fingerprinting (0.52 m vs. 1.3 m MAE) on spatially sparse data collected in a realistic industrial indoor environment.

Motivation & Objective

• Address the limitations of deep learning-based fingerprinting (CNN-FP), which require dense datasets and lack statistical observation models for fusion.
• Enable reliable position tracking in sparse data regimes common in real-world industrial environments with complex multipath propagation.
• Develop a likelihood-based observation model using Gaussian processes to support fusion with other sensors (e.g., TOF) at the observation level.
• Reduce data acquisition and maintenance costs by enabling accurate tracking with minimal, targeted fingerprinting data from high-multipath areas.
• Provide a scalable, compact, and reliable alternative to CNN-FP that maintains performance under data sparsity and environmental changes.

Proposed method

• Compress raw channel measurements (CMs) into spatial features using a trained autoencoder (AE) to extract compact, informative representations.
• Extract propagation-specific features (e.g., from multipath components) to enhance spatial sensitivity and model physical signal behavior.
• Train individual Gaussian processes (GPs) for each feature subset using distance-dependent covariance kernels to model observation likelihoods.
• Integrate the learned GP likelihoods with time-of-flight (TOF) ranges and a dynamics model within a particle filter for joint state estimation.
• Use stationary, distance-based kernels to implicitly model spatial correlation and reliability, enabling uncertainty-aware tracking.
• Apply a particle filter with resampling to fuse GP-based likelihoods, TOF measurements, and motion dynamics for robust, real-time tracking.

Experimental results

Research questions

• RQ1 Can a likelihood-based fingerprinting framework outperform CNN-based methods in spatially sparse indoor environments with complex multipath propagation?
• RQ2 How does the performance of a GP-based observation model degrade under data sparsity compared to deep learning-based fingerprinting?
• RQ3 Can the fusion of GP-estimated likelihoods with TOF and dynamics models improve tracking accuracy in low-line-of-sight conditions?
• RQ4 Does the use of compact, autoencoder-compressed features enable accurate tracking with minimal training data?
• RQ5 How does the framework maintain reliability and accuracy when environmental changes occur or data is missing in certain regions?

Key findings

• The proposed GPR-AE-S framework achieves a mean absolute error (MAE) of 0.52 m on a spatially sparse fingerprinting dataset, significantly outperforming CNN-FP (1.3 m MAE) in the same conditions.
• In areas with low line-of-sight coverage, the GPR-AE-S framework maintains a C95 error of 1.60 m, while CNN-FP shows a much larger performance drop due to lack of reliable observation modeling.
• The framework exhibits lower performance degradation (MED of 0.287 m) on sparse data compared to CNN-FP, demonstrating superior robustness to data sparsity.
• By modeling observation likelihoods statistically, the framework enables effective fusion with TOF and dynamics models in a particle filter, unlike CNN-FP which only fuses at the state level.
• The use of distance-dependent kernels in GPR allows the system to estimate reliability and uncertainty, enabling stable tracking even in large, mixed-propagation environments.
• The framework reduces data collection and maintenance costs by requiring only targeted, high-multipath area recordings, making it practical for industrial deployment.