奥迪自动驾驶数据集 A2D2

奥迪公司的研究人员在发布的论文 A2D2: Audi Autonomous Driving Dataset 中，公布了其大型自动驾驶数据集：A2D2，同时还配备了相关教程，并提供开放下载。数据集目标为推进计算机视觉、机器学习、自动驾驶的商用和学术研究。数据类型包含：RGB 图像，也包括对应的 3D 点云数据，记录的数据是时间同步的。标注类型包括：目标 3D 包围框，语义分割，实例分割以及从汽车总线提取的数据。数据规模：标注的非序列数据，41,227 帧，都含有语义分割标注和点云标签。其中含有前置摄像头视野内目标 3D 包围框标注 12,497 帧。另外，该库还包括 392,556 连续帧的无标注的传感器数据。图像中的车牌和人脸都进行了模糊化处理。总数据量很大，达到 2.3 TB。 这套数据集提供了摄像头、激光雷达和车辆总线数据，允许开发人员和研究人员探索多模态传感器融合方法。传感器套件包括六个摄像头和五个LiDAR单元，可提供完整的360度覆盖范围。 车辆配备了额外的硬件，用于记录来自传感器套件和车辆总线的数据。摄像机通过LVDS连接到嵌入式计算机，而LiDAR传感器通过1G以太网交换机连接。每个LiDAR传感器都连接到充当时钟的GNSS接收器。另一个GNSS时钟用作网关和嵌入式计算机的时间主机。总线网关通过1G以太网连接到嵌入式计算机。所有数据都存储在具有48 TB SSD存储的防撞网存储设备中，并可以通过10G以太网进行访问。 该数据主要来自德国街道，包含RGB图像，也包括对应的3D点云数据，记录的数据是时间同步的。目标3D包围框，语义分割，实例分割以及从汽车总线提取的数据。标注的非序列数据，41227帧，都含有语义分割标注和点云标签。其中含有前置摄像头视野内目标3D包围框标注12497帧。另外，该库还包括 392,556 连续帧的无标注的传感器数据。 此数据集的好处是，与KITTI，Waymo等不同，可以将其用于商业场景。 奥迪数据集的开发团队是由计算机科学家，工程师，物理学家和数学家组成的跨学科团队，致力于应对AI，机器人技术和自动驾驶技术中当今最大的挑战。在感知方面的内部工作促成了A2D2的创建，其公开发布反映了奥迪的开放，合作和社区精神的团队文化，共同推进世界计算机视觉、机器学习、自动驾驶的商用和学术研究。 A2D2数据集地址：https://www.a2d2.audi/a2d2/en.html

Researchers from Audi Corporation released their large-scale autonomous driving dataset, A2D2, alongside relevant tutorials and open download access in their published paper *A2D2: Audi Autonomous Driving Dataset*. The dataset aims to advance commercial and academic research in computer vision, machine learning, and autonomous driving. Data types include RGB images and corresponding 3D point cloud data, with all recorded data being time-synchronized. Annotation types include 3D object bounding boxes, semantic segmentation, instance segmentation, and data extracted from the vehicle bus.

Regarding data scale: There are 41,227 frames of annotated non-sequential data, all equipped with semantic segmentation annotations and point cloud labels. Among these, 12,497 frames have 3D object bounding box annotations within the field of view of the front-facing camera. Additionally, the dataset repository includes 392,556 consecutive frames of unannotated sensor data. License plates and human faces in the images have been anonymized via blurring. The total data volume reaches 2.3 TB.

This dataset provides camera, LiDAR, and vehicle bus data, enabling developers and researchers to explore multimodal sensor fusion methods. The sensor suite consists of six cameras and five LiDAR units, delivering full 360-degree coverage.

The vehicle is equipped with additional hardware for recording data from the sensor suite and vehicle bus. Cameras are connected to the embedded computer via LVDS, while LiDAR sensors are connected via a 1G Ethernet switch. Each LiDAR sensor is linked to a GNSS receiver that acts as a clock source. A second GNSS clock serves as the time host for the gateway and embedded computer. The bus gateway connects to the embedded computer via 1G Ethernet. All data is stored in a crash-protected storage device with 48 TB of SSD storage, and is accessible via 10G Ethernet.

The data is primarily collected from streets in Germany, including RGB images and corresponding 3D point cloud data, with all recorded data being time-synchronized. Annotation types include 3D object bounding boxes, semantic segmentation, instance segmentation, and data extracted from the vehicle bus. There are 41,227 frames of annotated non-sequential data, all equipped with semantic segmentation annotations and point cloud labels. Among these, 12,497 frames have 3D object bounding box annotations within the field of view of the front-facing camera. Additionally, the dataset repository includes 392,556 consecutive frames of unannotated sensor data.

A key advantage of this dataset is that, unlike KITTI, Waymo and other similar datasets, it can be used for commercial scenarios.

The development team of the A2D2 dataset is an interdisciplinary group composed of computer scientists, engineers, physicists, and mathematicians, dedicated to addressing the most significant current challenges in AI, robotics, and autonomous driving technology. Internal perception-related work led to the creation of A2D2, and its public release reflects Audi's team culture of openness, collaboration, and community spirit, with the goal of jointly advancing commercial and academic research in global computer vision, machine learning, and autonomous driving.

Official website of A2D2 dataset: https://www.a2d2.audi/a2d2/en.html

搜集汇总

背景与挑战

背景概述

奥迪自动驾驶数据集A2D2是一个大型多模态数据集，包含41,227帧标注的非序列数据（含语义分割和点云标签）和392,556帧连续无标注传感器数据，总数据量达2.3TB。该数据集特别提供了六摄像头和五激光雷达的360度覆盖数据，支持多模态传感器融合研究，并明确允许商业用途，区别于KITTI等学术数据集。