Data on Anomalous Traffic - Temple University (TU-DAT)

TU-DAT dataset is developed in order to improve the accuracy of accident detection in ITS. This dataset contains a diverse set of accident types, weather conditions, and videos collected in challenging environments, enhancing the self-adaptability of accident detection methods in a variety of traffic situations.

In order to collect the road accident or anomalous videos, we developed a crawler written in Python to scrape the accident videos from news reporting and documentary websites. In addition, we also searched YouTube videos for each type of anomaly using text search queries (with slight variations, e.g., "unexpected object on the road", "pedestrian accident," etc.). To ensure that our method applies to roadside edge devices, we use only footage and images from traffic CCTV cameras.

We have collected around 210 videos varying around 24-30 FPS of road accidents through these steps with 17255 accident keyframes and 505245 regular frames. Given the difficulty of obtaining real-world traffic videos to analyze aggressive driving, we adapted the game simulator to generate road traffic video data to simulate aggressive driving behaviors such as speeding, tailgating, weaving in and out of traffic, and running red lights. We gathered approximately 40 videos of positive examples and 25 videos of negative examples. The statistics of our TU-DAT dataset is shown in the below table:

Conditions #Frames Accident Types #Frames
Day Light 9796 Weaving thru traffic 2417
Night/low light 1487 X-section accidents 6566
Foggy 445 Tailgating 1452
Rainy 1281 Driving Maneuvers 305
Snowy 274 Rear-end accidents 1215
Camera too far 211 Pedestrian Accidents 447

The link to the dataset can be found at: GitHub