Argoverse Dataset 自动驾驶数据集

Argoverse数据集是由Argo AI、卡内基梅隆大学、佐治亚理工学院发布的用于支持自动驾驶汽车3D Tracking和Motion Forecasting研究的数据集。数据集包括两个部分：Argoverse 3D跟踪与Argoverse运动预测。 Argoverse中的数据来自Argo AI的自动驾驶测试车辆在迈阿密和匹兹堡（这两个美国城市面临不同的城市驾驶挑战和当地驾驶习惯）运行的地区的子集。包括跨不同季节，天气条件和一天中不同时间的传感器数据或“日志段”的记录，以提供广泛的实际驾驶场景。 总车道覆盖范围：迈阿密204线性公里，匹兹堡86线性公里。 Argoverse是第一个包含高精地图的数据集，它包含了290KM的带有几何形状和语义信息的高精地图数据。 Argoverse高精地图坐标系采用UTM坐标系，UTM将全球分成60个Zone，每个Zone经度方向跨越6度，在UTM基础之上，Argoverse又将坐标系相对于单个城市的固定坐标进行偏移，从而得到最终数据集的地图坐标系。 Argoverse特点： 1、具有113个场景的3D跟踪注释的一个数据集 2、一个数据集，包含从1000多个驾驶小时中提取的324,557条有趣的车辆轨迹 3、两张具有车道中心线，交通方向，地面高度等的高清（HD）地图 4、一个API，用于将地图数据与传感器信息连接起来 一、数据是如何收集的？ 我们使用与Argo AI自动驾驶技术完全集成的相同福特融合混合动力车队收集了所有数据。我们包括来自两个LiDAR传感器，七个环形摄像头和两个前置立体声摄像头的数据。所有传感器都安装在屋顶： 1.1 激光雷达： 2.2.1、2个屋顶式LiDAR传感器 2.2.2、重叠40°垂直视野 2.2.3、射程200m 2.2.4、平均而言，我们的LiDAR传感器在10 Hz时产生约107,000个点的点云。 1.2 特定城市与区域 我们使用特定于城市的坐标系进行车辆定位。我们结合了基于GPS的定位方法和基于传感器的定位方法，为每个时间戳都包含6自由度定位。 1.3 立体摄影机 七个高分辨率环形摄像头（1920 x 1200），以30 Hz的频率记录，并具有360°视场 以5 Hz采样的两个前视立体摄像机（2056 x 2464） 1.4 内外部校准 每次驾驶会话的传感器测量结果都存储在“日志”中。对于每个日志，我们提供LiDAR和所有九个摄像机的内部和外部校准数据 地图包含丰富的几何和语义元数据，可以更好地理解3D场景。从地面高度到下一个交叉点为止的剩余距离，我们的地图使研究人员能够探索高清地图在机器人感知中的潜力。 二、Argoverse地图 有三个不同的组件使我们的地图与众不同： 1、几何级车道矢量图 我们的语义矢量地图传达了有用的车道级别细节，例如车道中心线，交通方向和交叉路口注释。通过这些功能以及更多功能，用户可以探索交通流经我们测试区域中城市街道和复杂路口的多种方式，并获得每个场景前后的全面情况。 2、可驾驶区域的栅格化地图 我们的地图包含一米网格分辨率的二进制可驱动区域标签。可驾驶区域是车辆可能驾驶的区域（尽管不一定合法）。我们在3D跟踪中概述的跟踪注释延伸到可行驶区域之外五米。我们称这个更大的区域为“关注区域”。 3、具有实际高度的栅格化地图 我们的地图包括一米分辨率的实值地面高度。借助我们的地图工具，用户可以去除不平坦地面上的LiDAR返回物，从而更易于检测物体。 三、Argoverse 3D追踪 Argoverse 3D跟踪是113个日志段的集合，其中包含3D对象跟踪注释。这些日志段（我们称为“序列”）的长度从15秒到30秒不等，总共包含11052条轨道。 我们的训练和验证集中的每个序列都包含对所有物体的注释，这些物体位于我们称为“可驾驶区域”（车辆可以行驶的区域）5米以内的区域。 是什么使这个数据集脱颖而出？ 用户可以构建算法，以利用Argoverse的高清地图中的详细信息。例如，一种算法可以使用地图在LiDAR返回处执行地面移除操作，或基于车道方向约束车辆方向。 资料注解 Argoverse在可驱动区域上或附近的所有关注对象上均包含无定形态3D边界长方体。 “无模态”是指每个长方体的3D范围代表对象在3D空间中的空间范围-不仅是观察到的像素或观察到的LiDAR返回的范围，对于被遮挡的对象来说较小，而对于仅从观察到的对象来说是模糊的一张脸。 通过将长方体拟合到在整个跟踪序列中观察到的每个对象的LiDAR返回，将自动生成我们的非模态注释。如果对象的整个空间范围在一帧中模棱两可，则可以使用来自前一帧或后一帧的信息来约束形状。随着时间的推移，无定形长方体的大小是固定的。数据集中的一些对象会动态更改大小（例如，汽车开着门），并导致不完全的无峰长方体拟合。 为了创建模态长方体，我们在每个时间步确定属于每个对象的点。这些信息以及每个对象的方向都来自人工注释者。 我们为15个对象类提供了地面真相标签。这些类中的两个包括位于我们定义的键类别之外的静态和动态对象，它们分别称为ON_ROAD_OBSTACLE和OTHER_MOVER。这些对象类在Argoverse 3D跟踪中所有带注释的对象上的分布如下所示： 四、Argoverse运动预测 训练和验证运动预测模型的数据集 Argoverse运动预测是一个精心挑选的324,557个场景集合，每个场景5秒，用于训练和验证。每个场景都包含以10 Hz采样的每个跟踪对象的2D鸟瞰质心。 为了创建这个集合，我们从自动驾驶测试车队中筛选了1000多个小时的驾驶数据，以查找最具挑战性的细分市场-包括显示交叉路口的车辆，向左转或向右转弯的车辆以及改变车道的车辆。 In June 2019, we released the Argoverse datasets to coincide with the appearance of our publication, Argoverse: 3D Tracking with Forecasting and Rich Maps, in CVPR 2019. When referencing this publication or any of the materials we provide, please use the following citation:

Argoverse is a dataset released by Argo AI, Carnegie Mellon University, and Georgia Institute of Technology to support research on 3D Tracking and Motion Forecasting for autonomous vehicles. The dataset consists of two parts: Argoverse 3D Tracking and Argoverse Motion Forecasting.

The data in Argoverse comes from subsets of regions where Argo AI's autonomous test vehicles operated in Miami and Pittsburgh—two U.S. cities with distinct urban driving challenges and local driving habits. It includes records of sensor data, or "log segments", across different seasons, weather conditions, and times of day to cover a wide range of real-world driving scenarios. The total lane coverage is 204 linear kilometers in Miami and 86 linear kilometers in Pittsburgh. Argoverse is the first dataset to include high-definition maps, with 290 km of HD map data containing geometric and semantic information. The Argoverse HD map uses the UTM coordinate system, which divides the globe into 60 zones, each spanning 6 degrees of longitude. On top of the UTM framework, Argoverse offsets the coordinate system relative to the fixed coordinates of individual cities to obtain the final map coordinate system for the dataset.

Key features of Argoverse:
1. A dataset with 3D tracking annotations for 113 scenarios
2. A dataset containing 324,557 interesting vehicle trajectories extracted from over 1,000 driving hours
3. Two high-definition (HD) maps with lane centerlines, traffic directions, ground heights, and other relevant information
4. An API for connecting map data with sensor information

1. Data Collection
All data was collected using the same Ford Fusion Hybrid fleet fully integrated with Argo AI's autonomous driving technology. We include data from two LiDAR sensors, seven surround cameras, and two front-facing stereo cameras, all mounted on the roof:
1.1 LiDAR
1.1.1 Two roof-mounted LiDAR sensors
1.1.2 40° overlapping vertical field of view
1.1.3 200m range
1.1.4 On average, our LiDAR sensors generate point clouds of approximately 107,000 points at 10 Hz.
1.2 City-Specific Coordinate Systems for Vehicle Positioning
We use city-specific coordinate systems for vehicle positioning. We combine GPS-based positioning and sensor-based positioning to provide 6-degree-of-freedom positioning for each timestamp.
1.3 Stereo Cameras
Seven high-resolution surround cameras (1920 x 1200) record at 30 Hz with a 360° field of view. Two front-facing stereo cameras (2056 x 2464) are sampled at 5 Hz.
1.4 Intrinsic and Extrinsic Calibration
Sensor measurements from each driving session are stored in a "log". For each log, we provide intrinsic and extrinsic calibration data for the LiDAR and all nine cameras.

The map contains rich geometric and semantic metadata to enable better understanding of 3D scenes. From ground height to remaining distance to the next intersection, our map allows researchers to explore the potential of HD maps in robotic perception.

2. Argoverse Maps
Three distinct components set our map apart:
1. Geometric Lane Vector Map
Our semantic vector map conveys useful lane-level details such as lane centerlines, traffic directions, and intersection annotations. With these and other features, users can explore the multiple ways traffic flows through urban streets and complex intersections in our test regions, and gain a comprehensive view of the context before and after each scene.
2. Rasterized Map of Drivable Areas
Our map contains binary drivable area labels at 1-meter grid resolution. Drivable areas are regions where vehicles may drive (though not necessarily legally). The tracking annotations outlined in 3D Tracking extend five meters beyond the drivable area; we refer to this larger region as the "region of interest".
3. Rasterized Map with Actual Ground Height
Our map includes real-valued ground heights at 1-meter resolution. With our mapping tools, users can remove LiDAR returns from uneven terrain, making object detection easier.

3. Argoverse 3D Tracking
Argoverse 3D Tracking is a collection of 113 log segments containing 3D object tracking annotations. These log segments, which we refer to as "sequences", range in length from 15 to 30 seconds, totaling 11,052 tracks. Each sequence in our training and validation sets contains annotations for all objects within five meters of the "drivable area"—the region where vehicles can drive.

What makes this dataset stand out?
Users can build algorithms that leverage the detailed information in Argoverse's HD maps. For example, an algorithm could use the map to perform ground removal on LiDAR returns, or constrain vehicle orientation based on lane directions.

Data Annotations
Argoverse includes amodal 3D bounding cuboids for all objects of interest on or near the drivable area. "Amodal" means that the 3D extent of each cuboid represents the spatial extent of the object in 3D space—not just the range of observed pixels or observed LiDAR returns. For occluded objects, this is a more complete representation, while for objects visible only from a single face, it may be ambiguous. Our amodal annotations are automatically generated by fitting cuboids to the LiDAR returns of each object observed across the entire tracking sequence. If the full spatial extent of an object is ambiguous in one frame, information from previous or subsequent frames can be used to constrain the shape. The size of the amodal cuboid remains fixed over time. Some objects in the dataset dynamically change size (e.g., a car with an open door), leading to imperfect amodal cuboid fitting.
To create modal cuboids, we determine which points belong to each object at each timestep. This information, along with the orientation of each object, comes from human annotators.
We provide ground-truth labels for 15 object classes. Two of these classes include static and dynamic objects outside the key categories we defined, namely ON_ROAD_OBSTACLE and OTHER_MOVER. The distribution of these object classes across all annotated objects in Argoverse 3D Tracking is as follows:

4. Argoverse Motion Forecasting
Argoverse Motion Forecasting is a carefully curated collection of 324,557 scenarios, each 5 seconds long, for training and validation. Each scenario contains the 2D bird's-eye view centroid of each tracked object sampled at 10 Hz.
To create this collection, we screened over 1,000 hours of driving data from the autonomous test fleet to identify the most challenging segments—including vehicles at intersections, vehicles turning left or right, and vehicles changing lanes.

In June 2019, we released the Argoverse datasets to coincide with the publication of our paper, *Argoverse: 3D Tracking with Forecasting and Rich Maps*, at CVPR 2019. When referencing this publication or any of the materials we provide, please use the following citation:

搜集汇总

背景与挑战

背景概述

Argoverse Dataset是一个用于自动驾驶研究的综合性数据集，包含3D跟踪和运动预测两部分，数据来自迈阿密和匹兹堡的测试车辆，覆盖290公里的车道范围。其特点是包含高精地图和丰富的传感器数据（如LiDAR和摄像头），适用于复杂的城市驾驶场景研究。