POSNNET: Positioning Neural Network Estimation Tool

Many applications rely on IMU and GNSS data for localization, typically fused using algorithms such as the Kalman filter or its variants. Although this approach provides accurate localization without requiring additional infrastructure, its performance degrades in GNSS-denied environments. Deep learning methods have been proposed to address this issue in real time; however, because they do not leverage future data, they yield suboptimal reconstructions in applications that do not require real-time processing. To overcome this limitation, we introduce POSNNET, a post-processing framework designed to reconstruct position and velocity during prolonged periods of weak or absent GNSS signals. POSNNET is intended for localization scenarios in which real-time processing is not required, including sports performance analysis, infrastructure-free road mapping, and wildlife tracking, although its applicability across these domains requires further validation. The framework estimates velocity using state-of-the-art neural networks that leverage both past and future IMU and GNSS data. Position is subsequently recovered through weighted bidirectional integration. To improve reconstruction accuracy and consistency, the outputs of multiple neural networks are combined. POSNNET was validated offline against causal real-time baselines using simulated GNSS outages in a 100-h outdoor equestrian dataset characterized by higher speeds and more complex dynamics than typical pedestrian motion, with the sensor-fusion output serving as the reference trajectory. It achieved average trajectory errors, relative to this reference, of 4.07 m for simulated GNSS outages lasting from 30 s to 2 min and 7.06 m for outages lasting from 2 to 5 min, corresponding to an 93.4% (for outages lasting 30 s to 2 min) and 95.6% (for outages lasting 2 to 5 min) improvement in accuracy over the strongest evaluated real-time state-of-the-art method.

Authors

Institutions

Publication Details

Journal
Sensors
Published
2026-11-18
DOI
https://doi.org/10.3390/s26227282
Primary Topic
Inertial Sensor and Navigation
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

POSNNET: Positioning Neural Network Estimation Tool

Kévin Cédric Guyard, Jonathan Bertolaccini, Stéphane Montavon
6 links
Sensors
Inertial Sensor and Navigation
article

POSNNET: Positioning Neural Network Estimation Tool

Kévin Cédric Guyard, Jonathan Bertolaccini, Stéphane Montavon
article en

Abstract

Many applications rely on IMU and GNSS data for localization, typically fused using algorithms such as the Kalman filter or its variants. Although this approach provides accurate localization without requiring additional infrastructure, its performance degrades in GNSS-denied environments. Deep learning methods have been proposed to address this issue in real time; however, because they do not leverage future data, they yield suboptimal reconstructions in applications that do not require real-time processing. To overcome this limitation, we introduce POSNNET, a post-processing framework designed to reconstruct position and velocity during prolonged periods of weak or absent GNSS signals. POSNNET is intended for localization scenarios in which real-time processing is not required, including sports performance analysis, infrastructure-free road mapping, and wildlife tracking, although its applicability across these domains requires further validation. The framework estimates velocity using state-of-the-art neural networks that leverage both past and future IMU and GNSS data. Position is subsequently recovered through weighted bidirectional integration. To improve reconstruction accuracy and consistency, the outputs of multiple neural networks are combined. POSNNET was validated offline against causal real-time baselines using simulated GNSS outages in a 100-h outdoor equestrian dataset characterized by higher speeds and more complex dynamics than typical pedestrian motion, with the sensor-fusion output serving as the reference trajectory. It achieved average trajectory errors, relative to this reference, of 4.07 m for simulated GNSS outages lasting from 30 s to 2 min and 7.06 m for outages lasting from 2 to 5 min, corresponding to an 93.4% (for outages lasting 30 s to 2 min) and 95.6% (for outages lasting 2 to 5 min) improvement in accuracy over the strongest evaluated real-time state-of-the-art method.

SensorsVol. 26(22)
University of Geneva (CH), HES-SO University of Applied Sciences and Arts Western Switzerland (CH), HES-SO Genève (CH)
Innosuisse - Schweizerische Agentur für Innovationsförderung
Inertial Sensor and Navigation
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon