winter-driving-dataset
收藏资源简介:
AD冬季驾驶数据集是一个用于自动驾驶研究的开源多模态数据集,由AD Assurance Lab与西密歇根大学和MCity合作创建。该数据集专注于在恶劣冬季条件下的感知与物理建模,旨在为车道检测、物体检测等任务提供基准数据。数据集规模在10万到100万样本之间,包含图像分割和物体检测任务类别,适用于自动驾驶、车道检测、车辆动力学和天气遥测等研究领域。数据集采用CC BY 4.0许可证。数据集结构包括metadata目录(包含表格化元数据和注释)、models目录(提供预训练的车道检测模型检查点,支持ONNX和PyTorch格式)和images目录(包含mcity_wspi和reva_perception两个子集,其中mcity_wspi按动态序列组织独立压缩包,每个序列对应特定驾驶场景,reva_perception包含标注的训练图像集和未标注的序列帧)。每个驾驶序列中的相机帧以25Hz频率采集,并与车辆动力学数据(如CAN总线转向角、油门/刹车输入)、惯性响应数据(高精度6自由度IMU数据)、GPS/RTK地理坐标轨迹以及移动天气数据(如MARWIS道路条件分类)同步对齐。该数据集对应研究论文《MI-Snow1000: A Comprehensive Dataset and Benchmark for Lane Detection in Adverse Winter Conditions》,专注于冬季恶劣条件下的车道检测。
The AD Winter Driving Dataset is an open-source multimodal dataset for autonomous driving research, created by AD Assurance Lab in collaboration with Western Michigan University and MCity. It focuses on perception and physical modeling under harsh winter conditions, aiming to provide benchmark data for tasks such as lane detection and object detection. The dataset size ranges from 100,000 to 1,000,000 samples, includes image segmentation and object detection task categories, and is applicable to research areas like autonomous driving, lane detection, vehicle dynamics, and weather telemetry. It is licensed under CC BY 4.0. The dataset structure consists of three main parts: the metadata/ directory containing tabular metadata and annotations; the models/ directory providing pre-trained lane detection model checkpoints in ONNX and PyTorch formats; and the images/ directory containing image data, divided into mcity_wspi/ and reva_perception/ subsets. mcity_wspi/ includes independently compressed packages organized by dynamic sequences, each corresponding to specific driving scenarios; reva_perception/ contains annotated training image sets (in TuSimple format) and unlabeled sequence frames. Camera frames in each driving sequence are collected at 25Hz and synchronized with various physical variables, including vehicle dynamics data (CAN bus steering angle, throttle/brake inputs, brake torque, wheel speed), inertial response data (high-precision 6-DOF IMU rotation rates and linear acceleration, including vertical ride responses), GPS/RTK precise geographic coordinate trajectories, and mobile weather data (MARWIS real-time optical road condition classification and surface friction estimates). This dataset corresponds to the research paper MI-Snow1000: A Comprehensive Dataset and Benchmark for Lane Detection in Adverse Winter Conditions, focusing on providing a comprehensive dataset and benchmark for lane detection in adverse winter conditions.
AD Winter Driving Dataset 多模态感知与物理数据集
基本信息
- 数据集名称: AD Winter Driving Dataset
- 任务类型: 图像分割、目标检测
- 标签: 自动驾驶、车道检测、车辆动力学、天气遥测
- 许可协议: CC-BY-4.0
- 规模: 100K < n < 1M
- 创建机构: AD Assurance Lab(与西密歇根大学、MCity合作)
数据集结构
winter-driving-dataset/ ├── metadata/ # 表格元数据和标注(位于GitHub仓库) ├── models/ # 预训练车道检测模型 │ ├── onnx/ # ONNX float32/float16 格式检查点 │ └── pytorch/ # PyTorch .pt/.pth 检查点 └── images/ ├── mcity_wspi/ # 各动力学序列的独立zip包 └── reva_perception/ ├── labeled_set_images.zip # 标注的TuSimple格式训练图像 └── non_labeled_set/ # 未标注的序列帧
数据内容与特点
采集参数
- 相机帧率: 25Hz
- 对齐的物理变量:
- 车辆动力学:CAN总线转向角、油门/刹车输入、刹车扭矩、四个车轮转速
- 惯性响应:6自由度IMU旋转速率和线性加速度(包含垂直乘坐响应)
- GPS/RTK:精确的地理坐标轨迹
- 移动气象数据(MARWIS):实时光学路面分类和路面摩擦系数估计
图像数据
- 包含标注的TuSimple格式训练图像
- 包含未标注的连续帧序列
- 各动力学序列以独立zip包形式组织
预训练模型
- 提供车道检测的预训练检查点
- 支持ONNX和PyTorch两种格式
使用方法
下载特定运行图像
bash python3 tools/download_wspi.py --run jan27-downtown-1_images --dest_dir ./data
下载预训练车道模型
bash python3 tools/download_models.py --model all --dest_dir ./models
引用信息
bibtex @article{tye2026misnow1000, title={MI-Snow1000: A Comprehensive Dataset and Benchmark for Lane Detection in Adverse Winter Conditions}, author={Tye, Eugene and Clinton, Catherine and Asher, Zachary and Fong, Alvis}, journal={SAE International Journal of Connected and Autonomous Vehicles}, volume={9}, number={2}, pages={112--128}, year={2026}, publisher={SAE International} }
相关资源
- 主GitHub仓库(包含代码、同步CSV元数据、加载器和开发者文档):https://github.com/AD-Assurance-Lab/winter-driving-dataset




