A large-scale trajectory dataset of traffic participants are critical for the research areas, such as intention recognition, trajectory prediction, motion planning, representation learning and imitation learning, etc.
Both industry and academia are interested in high-quality trajectory datasets because many widely used approaches are data-driven and heavily rely on trajectory data from real traffic.
NGSIM dataset is so far the most popular dataset of naturalistic vehicle trajectories among the behavior-related research communities. The raw data of the dataset were collected by cameras mounted on tall buildings near the US-101 and I-80 highways of U.S. in a period from 2005 to 2006. Most of the extracted trajectory data are acceptable and utilizable. However, uncompleted trajectories, physically illogical vehicle speeds and accelerations existed in the dataset. Researchers have proposed various methods to rectify the errors, but it can only improve a small part of them. Then, HighD dataset was proposed in 2018. It was constructed by using a drone with high-resolution cameras and capturing the naturalistic traffic on German highways. By collecting data from a bird’s eye view, highD can provide more accurate vehicle locations and motions. Yet, HighD only contains straight road scenarios on highways, such as simple car following and lane change. On the other hand, both NGSIM (from U.S.) and HighD (from Germany) can not fully reflect the characteristics of Chinese traffic due to the differences of road conditions, traffic rules and driving styles in different countries.
For this reason, we publish a large-scale trajectory dataset of Chinese traffic participants called Mirror-Traffic, which is constructed based on abundant real traffic videos from both bird's eye view and road-side cameras.
The image processing and deep learning methods are used to identify and track the vehicles, and the extracted track is filtered by combining automatic and manual methods. The precision of trajectory can reach centimeter-level.
| duration | 30.00(min) |
| driven kilometers | 109.8(km) |
| driven hours | 1.58(h) |
| trajectories num | 760 |
| merge in num | 62 |
| duration | 30.00(min) |
| driven kilometers | 82.1(km) |
| driven hours | 1.13(h) |
| trajectories num | 556 |
| departure num | 290 |
| duration | 23.7(min) |
| driven kilometers | 62.5(km) |
| driven hours | 0.89(h) |
| trajectories num | 511 |
| merge in num | 175 |
The Mirror-Traffic dataset is only free for non-commercial use. If you are interested in commercial use, please contact with us via :
wangbaozong@tsari.tsinghua.edu.cn
