Finnish Winter Driving Dataset

This is a Winter Driving Dataset with high-quality GNSS/INS readings, forward-facing images, and 32-channel lidar scans (top mounted) for 1000 samples. Road segmentation annotations are provided for the images. The data is used in this paper. The source code for the paper is available in github. Image data has been anonymized so that it doesn't contain personal data. This work was funded by Henry Ford Foundation Finland and Aalto Doctoral School. The exact sensor setup is: GNSS/INS: Novatel PWRPAK 7DE2 Camera: FLIR blackfly S 2448x2048 Lidar: Velodyne VLP-32C The dataset is divided into countryside and suburb subsets, both containing a test sequence with 400 samples and a validation sequence with 100 samples. The file structure is as follows: finnish_winter_driving_dataset/├─ processed/│ ├─ countryside_test/│ │ ├─ data/│ │ │ ├─ images/│ │ │ │ ├─ 0.png│ │ │ │ ├─ ...│ │ │ ├─ scans/│ │ │ │ ├─ 0.npy│ │ │ │ ├─ ...│ │ │ ├─ pose_ids.txt│ │ │ ├─ poses.csv│ │ ├─ labels│ │ │ ├─ 0.png│ │ │ ├─ ...│ ├─ countryside_validation/│ │ ├─ ...│ ├─ suburb_test/│ │ ├─ ...│ ├─ suburb_validation/│ │ ├─ ... ├─ calib.yaml images folder contains undistorted images scans folder contains lidar scans separated into scan rings in numpy npy format. poses.csv contains latitude, longitude, and azimuth measurements at 10 Hz pose_ids.txt contains the id of each sample in poses.csv. The pose for the n:th sample can be retrieved by taking the n:th pose_id from pose_ids.txt and then taking the pose_id:th value from poses.csv. labels folder contains the manual road segmentation ground truth labels for the images calib.yaml contains the extrinsic and intrinsic calibration parameters. If you use this dataset, please cite @article{alamikkotervo2024trajectory, title={Trajectory-based Road Autolabeling with Lidar-Camera Fusion in Winter Conditions}, author={Alamikkotervo, Eerik and Toikka, Henrik and Tammi, Kari and Ojala, Risto}, journal={arXiv preprint arXiv:2412.02370}, year={2024} }本数据集为冬季驾驶数据集，包含1000组高质量全球导航卫星系统/惯性导航系统（GNSS/INS）观测数据、前视图像以及32线激光雷达扫描数据（车顶安装）。图像附带道路分割标注，相关内容已应用于本论文，论文的源代码可在GitHub平台获取。图像数据已完成匿名化处理，不包含个人敏感信息。本研究由芬兰亨利·福特基金会（Henry Ford Foundation Finland）与阿尔托大学博士研究生院（Aalto Doctoral School）资助。 具体传感器配置如下： GNSS/INS：诺瓦泰尔PWRPAK 7DE2（Novatel PWRPAK 7DE2） 相机：FLIR Blackfly S 2448×2048 激光雷达：Velodyne VLP-32C 本数据集分为乡村与郊区两个子集，每个子集均包含400组样本的测试序列与100组样本的验证序列。文件结构如下： finnish_winter_driving_dataset/ ├─ processed/ │ ├─ countryside_test/ │ │ ├─ data/ │ │ │ ├─ images/ │ │ │ │ ├─ 0.png │ │ │ │ ├─ ... │ │ │ ├─ scans/ │ │ │ │ ├─ 0.npy │ │ │ │ ├─ ... │ │ │ ├─ pose_ids.txt │ │ │ ├─ poses.csv │ │ ├─ labels/ │ │ │ ├─ 0.png │ │ │ ├─ ... │ ├─ countryside_validation/ │ │ ├─ ... │ ├─ suburb_test/ │ │ ├─ ... │ ├─ suburb_validation/ │ │ ├─ ... ├─ calib.yaml images文件夹存储已完成畸变校正的图像； scans文件夹存储以numpy的npy格式按扫描环分割的激光雷达扫描数据； poses.csv以10Hz的采样频率存储纬度、经度与方位角观测数据； pose_ids.txt存储poses.csv中每个样本的编号。第n个样本的位姿可通过以下方式获取：从pose_ids.txt中提取第n个样本对应的编号，再从poses.csv中调取该编号对应行的位姿数据； labels文件夹存储针对图像的人工标注道路分割真值标签； calib.yaml存储外参与内参校准参数。 若使用本数据集，请引用以下论文： @article{alamikkotervo2024trajectory, title={Trajectory-based Road Autolabeling with Lidar-Camera Fusion in Winter Conditions}, author={Alamikkotervo, Eerik and Toikka, Henrik and Tammi, Kari and Ojala, Risto}, journal={arXiv preprint arXiv:2412.02370}, year={2024} }