Warning message

Submissions for this form are closed.

Bosch Small Traffic Lights Dataset

Abstract

We present the Bosch Small Traffic Lights Dataset, an accurate dataset for vision-based traffic light detection. Vision-only based traffic light detection and tracking is a vital step on the way to fully automated driving in urban environments. We hope that this dataset allows for easy testing of objection detection approaches, especially for small objects in larger images.

The scenes cover a decent variety of road scenes and typical difficulties:
  • Busy street scenes inner-city
  • Suburban multilane roads with varying traffic density
  • Dense stop-and-go traffic
  • Road-works
  • Strong changes in illumination/exposure
  • Overcast sky with light rain
  • Flickering/Fluctuating traffic lights
  • Multiple visible traffic lights
  • Image parts that can be confused with traffic lights (e.g. large round tail lights)

Data description

This dataset contains 13427 camera images at a resolution of 1280x720 pixels and contains about 24000 annotated traffic lights. The annotations include bounding boxes of traffic lights as well as the current state (active light) of each traffic light.
The camera images are provided as raw 12bit HDR images taken with a red-clear-clear-blue filter and as reconstructed 8-bit RGB color images. The RGB images are provided for debugging and can also be used for training. However, the RGB conversion process has some drawbacks. Some of the converted images may contain artifacts and the color distribution may seem unusual.

Dataset specifications:

    Training set:
    • 5093 images
    • Annotated about every 2 seconds
    • 10756 annotated traffic lights
    • Median traffic lights width: ~8.6 pixels
    • 15 different labels
    • 170 lights are partially occluded
    Test set:
    • 8334 consecutive images
    • Annotated at about 15 fps
    • 13486 annotated traffic lights
    • Median traffic light width: 8.5 pixels
    • 4 labels (red, yellow, green, off)
    • 2088 lights are partially occluded
For the test set, every frame is annotated and temporal information was used to improve the label accuracy. The test-set was recorded independently from the training set, but within the same region. The data-set was created to prototype traffic light detection approaches, it is not intended to cover all cases and not to be used for production.

Example images:

References

The dataset has been created as part of our ICRA 2017 publication
A Deep Learning Approach to Traffic Lights: Detection, Tracking, and Classification (video)
If you publish work based on this data, please cite the following article:

@inproceedings{BehrendtNovak2017ICRA,
  title={A Deep Learning Approach to Traffic Lights: Detection, Tracking, and Classification},
  author={Behrendt, Karsten and Novak, Libor},
  booktitle={Robotics and Automation (ICRA), 2017 IEEE International Conference on},
  organization={IEEE}
}

Sample scripts

Sample scripts for reading the dataset are available at https://github.com/bosch-ros-pkg/bstld. Contributions are very welcome.

Acknowledgements

This work was conducted at the Bosch North America Research department, Palo Alto, California.

License

The dataset is released explicitly for non-commercial use only. The full license can be viewed here.
Additional data, such as unlabeled frames, odometry, and other vehicle information may be available for researchers on request.

Dataset Download

The dataset is available via Zenodo here: https://zenodo.org/doi/10.5281/zenodo.12706045.

Note: There is no password required for any of the zip archives. To extract the files download all parts of the zip files into the same folder and use 7-zip to extract the files.