Reliable vehicle localization remains challenging in GNSS-limited and GNSS-denied environments. This challenge becomes particularly severe under harsh Nordic winter conditions, where road markings, traffic signs, and visual landmarks are often obscured by snow. This paper presents SnowPole-GeoLoc, an open-source software framework for snow pole geo-localization using GNSS and LiDAR data fusion. In this framework, snow poles are treated as stable and machine-perceivable roadside infrastructure landmarks. The framework integrates deep learning-based snow pole detection from LiDAR-derived images with GNSS-assisted geolocalization. This combination enables the estimation of absolute pole locations in global map coordinates. The software provides modules for ROS bag processing, visualization, coordinate transformation, and quantitative evaluation against ground-truth pole locations. SnowPole-GeoLoc is evaluated using real-world data collected along Norwegian highways with a 128-channel LiDAR sensor and continuous GNSS measurements. The software is modular, reproducible, and publicly released with pretrained models, datasets, and environment specifications. It can be used as a standalone snow pole geo-localization tool or as a core sub-module within end-to-end vehicle localization pipelines designed for winter-degraded sensing conditions.

This study presents an extensive evaluation of YOLO object detection architectures for identifying snow poles in LiDAR-derived imagery under challenging Nordic conditions. Building on our prior Snow-Pole Detection dataset [1] and LiDAR-GNSS localization framework [2], we benchmark six YOLO models-YOLOv5s, YOLOv7-tiny, YOLOv8n, YOLOv9t, YOLOv10n, and YOLOv11n-across multiple input modalities. We assess single-channel modalities (Reflectance, Signal, Near-Infrared) and six pseudo-color combinations derived from these channels. Model performance is quantified using Precision, Recall, mAP@50, mAP@50-95, and GPU inference latency. To enable systematic comparison, we define a composite Rank Score combining accuracy and real-time performance. Results show YOLOv9t achieves the highest detection accuracy, while YOLOv11n offers the best balance between accuracy and inference speed, making it suitable for real-time applications. Pseudo-color combinations, especially those fusing Near-Infrared, Signal, and Reflectance, outperform single modalities and yield the highest Rank Scores. We recommend multimodal LiDAR configurations such as Combination 4 and Combination 5 to enhance detection robustness. All datasets, code, and trained models are publicly available to support reproducibility via our GitHub repository 4 and the Mendeley dataset archive 5 .

The SnowPole Detection dataset is a comprehensive collection of labeled LiDAR images, specifically designed for snow pole detection in road environments. This dataset was collected using a high-resolution OS2-128 LiDAR sensor mounted on an autonomous vehicle research platform, covering diverse environments such as mountainous, open, and forested areas. The SnowPole Detection dataset supports applications in computer vision, with a particular focus on snow pole detection and localization. he OS2-128 LiDAR sensor captures point clouds, which are processed using the Ouster SDK to generate 360-degree images in four modalities: Near-IR, Signal, Reflectivity, and Range. To enhance usability, color images were generated by assigning the first three modalities (Near-IR, Signal, and Reflectivity) to the blue, green, and red channels, respectively, excluding the Range modality. Initial labeling was conducted using Roboflow, with further refinement in CVAT, resulting in high-quality annotations. The dataset comprises a total of 1,954 manually labeled images, divided into 1,367 training images, 390 validation images, and 197 test images, following a 70/20/10 split. Since the images across all modalities are pixel-aligned, the labels for the color images are also applicable to each modality individually. This structure allows researchers to directly use the dataset for snow pole detection tasks, whether focusing on color or individual LiDAR modalities. The SnowPole Detection dataset is publicly available at Mendeley. 

The image fusion algorithm based on deep learning possesses strong feature extraction capabilities and generalization. However, due to the uninterpretability of features in deep learning, the design of fusion strategies becomes quite challenging. To address this issue, we propose a two-stage training feature adaptive fusion network based on the VGG-19 network. We introduce a parallel cross-modal channel perception module to achieve more targeted feature fusion by capturing channel differences between different modal domains. At the same time, in order to enhance the preservation of salient features, we designed a dynamic multi-level spatial attention guidance module that utilizes the saliency information of deep features from the source image to guide the adaptive fusion of shallow features. Additionally, we propose a double inner-loop feature mutual information loss that enforces the correlation of modal information, promoting efficient convergence of the perception module and guidance module. This method not only preserves the unique features of each modal domain but also effectively integrates information across modal domains, improving the quality of image fusion. Finally, we also perform objective and subjective experiments on MSRS and TNO datasets, and analyze the method. Experiments show that the proposed method achieves superior performance in image fusion tasks, and its potential value in practical applications is verified. The source code will be publicly available at https://github.com/YQ-097/TransFusion

This study introduces a robust vehicle localization framework designed for GNSS-limited environments. The proposed approach dynamically integrates georeferenced snow poles—fixed markers used to delineate road boundaries in winter with LiDAR-based odometry to enhance vehicle positioning and navigation. By alternating between GNSS data and LiDAR-based localization depending on GNSS signal availability, the framework addresses the challenges of GNSS-denied environments while leveraging sparse GNSS signals when available. A newly developed dataset of 360-degree snow pole images, captured using an Ouster OS2-128 LiDAR sensor, demonstrates the system’s applicability for autonomous driving. The method achieves a median localization error of 8.39m in GNSS-denied conditions, significantly outperforming techniques like FastReg ( 35.68m ), and progressively improves to sub-meter accuracy as GNSS availability increases. The open-source pipeline, to be made available on https://github.com/bdps1989/Snow-pole-based-vehicle-localization, offers a scalable, reliable, and near real-time solution for autonomous navigation in Nordic winter conditions, advancing research in localization under adverse environments.