Tsinghua AIR Research Institute releases the world's first vehicle-road collaboration dataset DAIR-V2X

According to the Institute of Intelligent Industry of Tsinghua University, yesterday, at the 2021 World Intelligent Connected Vehicle Conference, the world's first vehicle-road collaborative DAIR-V2X dataset was officially released.

It is reported that this is the world's first vehicle-road collaboration dataset based on real road scenarios. The dataset was jointly released by Tsinghua University's Intelligent Industry Research Institute (AIR), Beijing High-level Autonomous Driving Demonstration Zone, and Beijing AutoNet Technology Development Co., Ltd., with technical support from Baidu Apollo.

IT Home learned that the DAIR-V2X vehicle-road collaborative dataset has three main characteristics: the first vehicle-road collaborative dataset, rich sensors, large data volume and high quality.

Specifically, the data set was collected from 10 kilometers of urban roads, 28 intersections, and 10 kilometers of highways in Beijing's high-level autonomous driving demonstration zone, including 72,890 frames of images with 3D annotations and 72,890 frames of point cloud data; the data sources include sensors such as vehicle-side cameras, vehicle-side lidars, road-side cameras, and road-side lidars; the data type includes rich 3D annotation information, an average of 20 3D frames/frame data, as well as rich scenes such as sunny/rainy days, daytime/nighttime, urban roads/highways (trajectory prediction, traffic flow indicator statistics, etc. can be further expanded in the future). The overall coverage includes three parts: single-vehicle-side + single-road-side multimodal data, vehicle-road collaborative multimodal data in the same time and space, and 2D&3D annotations corresponding to the data set.

Tsinghua University Institute of Intelligent Industry said that compared with data sets that only include single-vehicle or road-side data sets, the DAIR-V2X data set for vehicle-road collaboration released this time has overcome the previous problem of vehicle-road collaboration detection in the same time and space but data is not synchronized for the first time, and proposed a multi-modal fusion method for vehicle-road collaboration and provided detection indicators, solving the pain point of the vehicle-road collaboration industry in the past lack of real road scene data. Based on this data set, researchers can carry out related research on single-vehicle 3D detection tasks (including monocular, point cloud and multimodal forms), single-road 3D detection tasks (including monocular, point cloud and multimodal forms), vehicle-road collaboration perception 3D detection (including monocular, point cloud and multimodal and other combinations).