China launches V2X-Seq dataset for cooperative V2X research

Beijing (Gasgoo)- On May 11, the Beijing High-level Automated Driving Demonstration Area (BJHAD), in collaboration with Tsinghua University’s Institute for AI Industry Research (AIR), Baidu Apollo, the Beijing Academy of Artificial Intelligence (BAAI), and Beijing Car Network Technology Development Co., Ltd., launched the V2X-Seq, the world's first real-road-scenario-based time-sequence V2X dataset. 

Photo credit: BJHAD

The dataset is now available for download and usage by users in China, filling a crucial gap in the industry by accelerating the research on V2X time-sequence perception and cooperative trajectory prediction.

The second release of the V2X-Seq consists of two components, namely, the V2X time-sequence perception dataset and the V2X cooperative trajectory prediction dataset. Leveraging self-developed spatiotemporal trajectory matching algorithms, the V2X-Seq provides high-quality ground truth cooperative trajectory data. 

Additionally, based on the V2X-Seq dataset, three research tasks related to V2X 3D tracking and cooperative trajectory prediction have been defined, accompanied by a rich set of algorithm benchmarks. The dataset and its associated algorithms have been officially accepted by the CVPR 2023.

The V2X time-sequence perception dataset includes over 15,000 pairs of time-sequence images and point cloud data from cooperative perspectives of both vehicles and infrastructure. It offers comprehensive 3D object and Tracking ID annotations to support V2X 3D detection and tracking tasks.

The V2X cooperative trajectory prediction dataset covers 28 real-road intersections, comprising over 200,000 segments, including target trajectories and traffic light signals. Among these segments, approximately 50,000 are from cooperative perspectives of vehicles and infrastructure, with each segment lasting 10 seconds. The dataset provides trajectory information for eight common obstacle targets at a frequency of 10 Hz.