HDSVT: High-Density Semantic Vehicle Trajectory Dataset Based on a Cosmopolitan City Bridge

Trajectory prediction is crucial for autonomous driving, necessitating robust models supported by comprehensive datasets. Most existing vehicle trajectory datasets often lack long-duration trajectories or high vehicle density, limiting their use in complex scenarios. In this paper, we construct a High-Density Semantic Vehicle Trajectory Dataset (HDSVT), collected via Unmanned Aerial Vehicles (UAVs) over the Guangzhou Bridge. Our dataset is characterized by higher vehicle density, extended trajectory lengths, and semantic enhancements, which facilitate in-depth vehicle trajectory analysis. We provide its original UAV videos, after processing these videos, our dataset is divided into three parts: (i) pixel coordinates of vehicle trajectory and lane lines in each video, (ii) corresponding geographic coordinates and (iii) semantic promotion for trajectories and motions. In this high-density, long-range trajectory context, the dataset captures diverse driving behaviors and complex multi-vehicle interaction. With its rich data on diverse driving behaviors and complex multi-vehicle interactions, this dataset is not only suitable for trajectory and motion prediction but also serves broader applications in driving decision-making, traffic management, and long-term tracking of small objects.