VSLAB

Anticipating Accidents in Dashcam Videos

   

dataset

 Fig.1 

  A dashcam is a cheap aftermarket camera, which can be mounted inside a vehicle to record street-level visual observation from the driver's point-of-view (see Fig.1-Top-Right-Corner). In certain places such as Russia and Taiwan, dashcams are equipped on almost all new cars in the last three years. Hence, a large number of dashcam videos have been shared on video sharing websites such as YouTube.

  In particular, we target at accident videos with human annotated address information or GPS locations. In this way, we have collected various accident videos with high video quality (720p in resolution). The dataset consists of 620 videos captured in six major cities in Taiwan. Our diverse accidents include: 42.6% motorbike hits car, 19.7% car hits car, 15.6% motorbike hits motorbike, and 20% other type. Figure1 shows a few sample videos and their corresponding locations on Google map.     

  Our videos are more challenging than videos in the KITTI dataset due to the following reasons,

  • Complicated road scene: The street signs and billboards in Taiwan are significantly more complex than those in Europe.

  • Crowded streets: The number of moving cars, motorbikes, and pedestrians per frame are typically larger than other datasets.

  • Diverse accidents: Accidents involving cars, motorbikes, etc. are all included in our dataset.

Youtube Example 

Dataset [Download] [Evaluation code]

 

We manually annotate the temporal locations of accidents. Among the remaining 620 videos, we sample 1750 clips, where each clip consists of 100 frames (5 seconds). These clips contain 620 positive clips containing the moment of accident at the last 10 frames, and 1130 negative clips containing no accidents. We randomly split the dataset into training and testing, where the number of training clips is about three times the number of testing clips: 1284 training clips and 466 testing clips. 

 

 

Dataset Distribution

  Positive examples Negative examples Total
Training set  455 829 1284
Testing set 165 301 466
Total 620 1130 1730

 

 

Positive example

Negative example

 

Visualization

 

Publications

  • Fu-Hsiang Chan, Yu-Ting Chen, Yu Xiang, Min Sun, "Anticipating Accidents in Dashcam Videos ." ACCV 2016 Oral (pdf) (github).

Contact : Fu-Hsiang Chan (corgi1205@gmail.com)

Last update : March 20th, 2017