AIOdrive

全传感器套件 (3倍激光雷达、1倍SPAD激光雷达、4倍雷达、5倍RGB、5倍深度相机、IMU、全球定位系统)
高密度远距离激光雷达点云
来自SPAD-LiDAR的多回波点云
每个具有1000帧 (100 s) 的100序列
不包括车祸和违反交通规则在内的分布数据
用于5个摄像机视点的500,000注释图像
每个LiDAR/雷达传感器的100,000个带注释的帧
26M 2D/3D边界框精确注释为4个对象类别 (汽车、骑自行车的人、摩托车、行人)
跨时间注释对象身份以形成轨迹
对象属性，如截断/遮挡百分比、角和线速度、加速度、制动、转向、油门
顺序点云泛光分割: 为所有序列中的23个语义类注释的所有点; 属于前景对象的点也为唯一实例类注释。
视频全景分割: 为所有视频中的23个语义类注释的所有像素。属于前景对象的像素也被注释为一个唯一的实例类。
可免费用于非商业和商业用途

Full sensor suite (3x LiDAR, 1x SPAD LiDAR, 4x radar, 5x RGB cameras, 5x depth cameras, IMU, GPS)
High-density long-range LiDAR point clouds
Multi-echo point clouds from SPAD-LiDAR
100 sequences, each containing 1000 frames with a total duration of 100 seconds
Distributed data excluding traffic accidents and traffic law violations
500,000 annotated images captured from 5 camera viewpoints
100,000 annotated frames per LiDAR/radar sensor
26M precisely annotated 2D/3D bounding boxes across 4 object categories: car, cyclist, motorcycle, pedestrian
Cross-temporally annotated object identities to form complete trajectories
Object attributes including truncation/occlusion percentage, angular velocity, linear velocity, acceleration, braking status, steering angle, and throttle input
Sequential point cloud panoptic segmentation: All points are annotated with 23 semantic classes across all sequences; points belonging to foreground objects are also annotated with unique instance classes
Video panoptic segmentation: All pixels are annotated with 23 semantic classes across all videos; pixels belonging to foreground objects are also annotated with unique instance classes
Freely available for both non-commercial and commercial purposes

搜集汇总

背景与挑战

背景概述

AIOdrive是一个大规模自动驾驶综合感知数据集，包含多种传感器数据（如激光雷达、雷达、相机等）和丰富的注释信息（如2D/3D边界框、对象属性等），数据规模达2926.6GB，可免费用于非商业和商业用途。该数据集由卡内基梅隆大学于2021年发布，旨在支持自动驾驶相关研究。