*Important notice: This news reports on an unedited version of an accepted paper and is awaiting final editing. Therefore, the paper should not be regarded as conclusive or treated as established information.
Researchers have developed a geographic information system (GIS) that uses incremental motion estimates to advance a robotic vehicle’s position along a route when global navigation satellite system (GNSS) positioning is unavailable. Their landmark-aided, route-constrained localization framework was presented in Scientific Reports.
Study: GIS and landmark-aided autonomous vehicle localization for navigation continuity in urban corridors. Image Credit: Danko Creative/Shutterstock.com
Limitations of Existing Localization Approaches
GNSS provides the primary absolute positioning source for autonomous and connected vehicles. However, in urban corridors, its signals can
degrade through blockage, multipath, non-line-of-sight reception, interference, jamming, and spoofing.
Existing fallback approaches each carry limitations. For example, inertial methods accumulate error without external corrections; signal-of-opportunity methods depend on external radio sources; vision- and simultaneous localization and mapping (SLAM)-based methods require perception and data association; and high-definition-map methods demand dense map creation, maintenance, and repeated matching.
Sparse landmark methods reduce map burden but make reliable association critical. To address this, this research presented a GIS- and landmark-aided route-constrained framework that estimates one-dimensional route progress along a known route using incremental motion estimates, corrected by mapped intersections and parking lot access points.
Framework for Route-Constrained Vehicle Localization
The framework maintains route-level localization continuity when GNSS position is unavailable by combining three elements, namely, a GIS route constraint, incremental motion estimation, and sparse landmark correction.
A trusted initial position and destination define the planned route, incremental-distance estimates advance a scalar route-progress state, and mapped landmarks reset that progress when a valid event is accepted. The final latitude-longitude output comes from route geometry rather than independent two-dimensional estimation.
The route is built from OpenStreetMap road-centerline data using the Open-Source Routing Machine, and because the vehicle travels within a lane, the route is not treated as lane-level or survey-grade accurate.
Incremental distance is estimated from onboard motion data using two methods. The long short-term memory (LSTM) maps seven synchronized motion inputs to incremental distance, trained and tested on separate experimental runs, with five fits using different random seeds.
The Kalman Filter recursively combines vehicle dynamics with velocity and acceleration measurements, using a fixed parameterization without systematic sensitivity analysis.
Download a free copy of this page here!
Landmark correction involves detecting, associating, and correcting route progress. The experimental database contained 23 landmarks. Automatic detection was not evaluated, with predefined landmark-passage events used instead. A stationary threshold prevented progress updates during stops.
Evaluation used real-world data from a 2.53 km urban corridor, with a NovAtel GNSS/inertial system providing the offline reference trajectory and BESTGNSSVEL supplying recorded speed. Component ablation isolated the route constraint and landmark correction, and controlled simulations tested GNSS position outage and error.
Comparing Motion Estimators and Localization Performance
The evaluation moved from estimator-level performance to complete route-constrained localization. Across five independently initialized LSTM models, the independent test error was 0.0213 ± 0.003 m.
The Kalman Filter achieved lower incremental-distance error on every reported metric, with 0.0148 m. Training-set size did not explain this difference: increasing training samples did not monotonically reduce test error, and adding same-direction data did not improve results. The single-corridor dataset therefore does not support broader generalization claims.
Despite the estimator-level gap, final map-referenced localization was comparable. The LSTM produced 3.410 ± 0.126 m mean error versus 3.445 m for the Kalman Filter, with a correlation of 0.960 between error profiles. The estimator difference was more than two orders of magnitude smaller than the final error.
Component ablation showed unconstrained dead reckoning at 32.631 m mean error, falling to 22.059 m with the route constraint alone, 6.054 m with landmark correction alone, and 3.445 m with both. The route constraint keeps position on the roadway, while landmark resets limit accumulated progress error.
Landmark-event testing showed missed corrections caused the mildest degradation. Mean error rose from 3.445 m to between 3.731 and 5.689 m when 10% to 50% of corrections were missed, but to between 6.681 and 19.604 m when every correction was delayed by between half a second and two seconds. Misassociating a landmark produced the worst individual failures, with the worst case reaching a mean error of 66.933 m.
Under controlled GNSS outages, the fallback maintained a 2.723 m mean error, whereas holding the last valid position produced 242.800 m. Automatic landmark detection was not evaluated, and the experiment demonstrates localization without GNSS position updates rather than complete GNSS independence.
Practical Navigation Continuity in Urban Corridors
This study developed and evaluated a GIS- and landmark-aided, route-constrained localization framework designed to maintain navigation continuity when GNSS positioning is unavailable.
The framework advances vehicle progress along a known GIS route using incremental motion estimates, with sparse mapped landmarks correcting accumulated route-progress error.
GNSS position updates were excluded throughout the experimental corridor, though BESTGNSSVEL remained the recorded speed input, so the evaluation demonstrates continuity without GNSS position updates rather than complete GNSS independence. The Kalman Filter achieved lower incremental-distance error, yet both estimators supported comparable route-level localization once embedded in the complete framework.
Component ablation confirmed the value of combining route constraints with landmark corrections, and landmark-event perturbations showed that performance is more sensitive to delayed and misassociated corrections than to missed ones. The results demonstrate that lightweight route constraints and sparse landmark corrections can provide practical meter-level navigation continuity in urban corridors.
Journal Reference
Dasgupta, S., et al. (2026). GIS and landmark-aided autonomous vehicle localization for navigation continuity in urban corridors. Scientific Reports. DOI:10.1038/s41598-026-70368-x. https://www.nature.com/articles/s41598-026-70368-x.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.