The new release, described in the journal SoftwareX by Durga Prasad Bavirisetti of Dalarna University in Sweden, closes a conceptual loop that began with version 1.0. That earlier release provided tools for detecting snow poles from LiDAR point clouds—dense, three-dimensional scans of the environment produced by laser ranging sensors—and for estimating their absolute geographic positions by fusing those LiDAR observations with Global Navigation Satellite System (GNSS) data. The result was a set of georeferenced snow-pole landmarks, essentially a map of persistent, infrastructure-based reference points embedded in the winter landscape. But the original framework stopped there, identifying the landmark generation as a reusable sub-module for a broader localization system yet to come. Version 2.0 delivers that system, shifting the core functionality from landmark mapping to landmark-assisted vehicle localization.

The problem this addresses is one that autonomous driving researchers know well. GNSS signals—the constellation of satellite transmissions that underpin GPS and comparable systems—are remarkably useful in open conditions, but they degrade badly in urban canyons between tall buildings, under dense tree canopies, in heavy precipitation, and in the deep-cut road corridors of mountainous Nordic terrain. When a vehicle loses reliable satellite positioning, it must fall back on dead reckoning: estimating its own motion from onboard sensors. For LiDAR-equipped vehicles, this typically means registering consecutive point clouds against one another—finding the transformation that best aligns each new scan with the previous one—to infer how the vehicle has moved between frames. The technique works, but it suffers from a fundamental weakness: error accumulates. Each registration is imperfect, and those small errors compound over time, causing the estimated trajectory to drift away from the vehicle’s true path.

Version 2.0 of SnowPole-GeoLoc confronts this drift directly by exploiting the one thing a snowy Nordic road offers in abundance: poles. The framework’s new modules propagate vehicle motion from sequential LiDAR observations, associate locally observed poles with their counterparts in the georeferenced landmark map, and use these persistent landmarks to constrain the accumulated trajectory error. Conceptually, the approach resembles the way mariners once fixed their position from known coastal beacons. A vehicle driving through a GNSS-denied stretch of road can still detect snow poles with its LiDAR; because each pole’s absolute coordinates are already known from the landmark map generated during the geo-localization phase, matching an observed pole to a mapped pole provides an immediate, absolute positional constraint that snaps the drifting trajectory back into alignment with reality.

Under the hood, the software relies on FastReg, a fast and robust point-cloud registration method originally described in IEEE Robotics and Automation Letters, as its registration backend. Registration is the computational heart of incremental LiDAR navigation: given two partially overlapping point clouds, the algorithm must estimate the rigid transformation—translation and rotation—that best aligns them. FastReg was selected for its speed and resilience to partial overlap, both critical in real-time driving scenarios where scans may capture only fragments of the previous scene. Importantly, the developers have kept the architecture modular, so that alternative registration methods can be substituted without restructuring the overall workflow. This design choice reflects a philosophy of reproducibility and extensibility that runs through the entire project.

The release is comprehensive in what it ships. The repository contains the main end-to-end vehicle-localization script, a dedicated variant for evaluating localization performance at different GNSS-availability percentages, reusable geospatial and data-association utilities, ROS bag processing and visualization tools, pretrained model assets for snow-pole detection, georeferenced ground-truth poles, and example outputs from incremental navigation runs. A Conda environment specification pins the full software stack—Python 3.9.18, PyTorch, OpenCV, Open3D, and supporting libraries—so that other researchers can reproduce the results on Ubuntu 22.04.5 LTS without fighting dependency conflicts. The code is distributed under the permissive BSD-3-Clause license on GitHub, and the accompanying real-world datasets are publicly available through a dedicated Data in Brief publication and a Kaggle dataset.

Those datasets deserve particular attention, because they ground the software in genuine Nordic winter conditions rather than idealized laboratory scenarios. The SnowPole detection dataset, published in Data in Brief, provides comprehensive LiDAR imaging data collected specifically for detection and localization tasks in Nordic winter environments—conditions characterized not only by snow-covered poles but by the optical and geometric challenges that snow, frost, and darkness impose on laser sensors. The companion Kaggle dataset supplies the data underlying the snow-pole-based vehicle localization experiments. By pairing open software with open data, the project lowers the barrier to entry for research groups anywhere in the world who want to study navigation in GPS-degraded environments, a problem relevant far beyond Scandinavia.

The methodological foundations of version 2.0 were developed in a body of peer-reviewed work led by Bavirisetti and colleagues at the Norwegian University of Science and Technology (NTNU) and SINTEF Community in Trondheim. Most notably, a 2025 paper in IEEE Transactions on Intelligent Transportation Systems presented the vehicle localization framework using georeferenced snow poles and LiDAR in GNSS-limited environments under Nordic conditions, reporting trajectory-scale experimental results. An earlier contribution at the 27th International Conference on Information Fusion (FUSION) in 2024 laid out the pole detection and geospatial localization framework using LiDAR–GNSS data fusion. The contribution of the new SoftwareX article is deliberately distinct from these methodological papers: it is the open and reproducible integration of those localization capabilities into the previously published SnowPole-GeoLoc framework, transforming research code into a documented, licensable, community-accessible tool.

The significance of this work extends well beyond the immediate technical achievement. Snow poles are an unusually attractive class of infrastructure landmark for several reasons. They are already installed along vast stretches of northern roads at regular intervals, maintained by road authorities, and designed from the outset to be visually and physically conspicuous against a snow-covered background. They are passive—requiring no power, no communication, and no maintenance of embedded electronics—and they persist reliably through the very seasons when GNSS performance is often at its worst. In effect, version 2.0 converts existing passive road furniture into a cooperative positioning infrastructure, without the cost of deploying new transponders, radio beacons, or dedicated navigation markers. This is infrastructure-aware engineering in the truest sense: extracting maximal navigational value from what is already standing by the roadside.

The work is embedded in a broader research context focused on machine-sensible infrastructure for Nordic conditions, conducted as part of the project Machine Sensible Infrastructure under Nordic Conditions (Project No. 333875), with additional support from the Knowledge Foundation (KKS) in Sweden under the project Digital Image Analysis and Artificial Intelligence (Contract No. KKS-20230085). The framing is telling: rather than asking vehicles to cope alone with hostile environments, the research envisions roads themselves becoming legible to machine perception, with existing physical features serving as dependable references for sensors and algorithms. Snow poles, in this vision, become nodes in a distributed, passive positioning network.

For the autonomous driving community, the configurable GNSS-availability evaluation included in version 2.0 is likely to prove especially valuable. The tool allows researchers to systematically degrade or remove GNSS input—simulating partial signal loss or complete denial—and measure how the landmark-assisted localization system responds. This kind of controlled stress-testing is essential for safety-critical systems, where the performance envelope under sensor degradation matters as much as performance under nominal conditions. The ability to quantify trajectory drift and its correction across varying GNSS availability provides a template for evaluating other infrastructure-assisted navigation schemes as well.

What began as an effort simply to find snow poles in LiDAR scans has matured into a complete navigation paradigm for one of the most challenging driving environments on Earth. Version 2.0 of SnowPole-GeoLoc demonstrates that the answer to navigation failure in snowbound landscapes may not require new satellites, new transmitters, or new roadside electronics—only new eyes, in the form of LiDAR, and a way to recognize the humble poles that have guided winter drivers for decades. The framework is available on GitHub, and the researchers welcome questions at the contact address listed in the repository documentation. As autonomous systems venture further into regions where the sky’s signals fade and the ground turns white, infrastructure landmarks like the snow pole may prove to be exactly the anchor that machine navigation has been missing.