自动驾驶数据集NuScenes

nuScenes数据集 是自动驾驶公司nuTonomy建立的大规模自动驾驶数据集，该数据集不仅包含了Camera和Lidar，还记录了雷达数据，是我所知道的唯一一个有雷达数据的数据集。 这个数据集由1000个场景组成（即scenes，这就是该数据集名字的由来），每个scenes长度为20秒，包含了各种各样的情景。在每一个scenes中，有40个关键帧（key frames），也就是每秒钟有2个关键帧，其他的帧为sweeps。关键帧经过手工的标注，每一帧中都有了若干个annotation，标注的形式为bounding box。不仅标注了大小、范围、还有类别、可见程度等等。这个数据集不久前发布了一个teaser版本（包含100个scenes），正式版（1000个scenes）的数据要2019年发布。 这个数据集在sample的数量上、标注的形式上都非常好，记录了车的自身运动轨迹（相对于全局坐标），包含了非常多的传感器，可以用来实现更加智慧的识别算法和感知融合算法。

The nuScenes dataset is a large-scale autonomous driving dataset developed by autonomous driving firm nuTonomy. It encompasses data from cameras, LiDAR, and radar sensors, making it the only dataset I am aware of that includes radar data. This dataset comprises 1000 scenes (which is the origin of its name), each with a duration of 20 seconds and covering a diverse range of driving scenarios. Within each scene, there are 40 key frames, equivalent to 2 key frames per second, with the remaining frames categorized as sweeps. The key frames are manually annotated, with each frame containing multiple bounding box annotations that record attributes such as object size, spatial extent, category, visibility, and other relevant details. A teaser version of the dataset (containing 100 scenes) was released recently, while the full version (with 1000 scenes) was scheduled for release in 2019. This dataset excels in both the number of samples and the quality of its annotation framework, as it records the ego-motion trajectory of the autonomous vehicle relative to the global coordinate system, integrates a rich array of sensor modalities, and can be used to develop more advanced recognition and perception fusion algorithms.

搜集汇总

背景与挑战

背景概述

NuScenes是由nuTonomy公司建立的大规模自动驾驶数据集，包含Camera、Lidar和雷达数据，由1000个场景组成，每个场景20秒，关键帧经过手工标注，适用于高级识别和感知融合算法开发。