This repository accompanies the Austrian Highway Traffic Sign Data Set (ATSD), a large data set of annotated traffic scene and -sign images that can be used for training and benchmarking traffic sign recognition models. The repository contains example Jupyter notebooks with Python code for loading images and meta data, augmenting the data set, and applying trained traffic sign detection- and classification models.

Clone the repository and create a new environment from requirements.txt:

# pip
pip install virtualenv
python -m virtualenv <env_name>
<env_name>\Scripts\activate
pip install -r requirements.txt

# conda
conda create --name <env_name> --file requirements.txt

The main contents of this repository are Jupyter notebooks for illustrating how to work effectively with the data set, as well as trained baseline detection- and classification models. The data set itself can be downloaded here.

• Classification_Preparation.ipynb shows how traffic sign images and related meta-data from ATSD-Signs can be loaded, analyzed and augmented with traffic-sign-specific augmentation methods. It forms the basis for subsequently training classification models, although the model training itself is not covered.
• Detection_Preparation.ipynb shows how traffic scene images and related meta-data from ATSD-Scenes can be loaded and converted to a YOLO/Darknet compatible format. It forms the basis for subsequently training detection models, although the model training itself is not covered.
• Classification.ipynb shows how trained traffic sign classifiers can be applied to and evaluated on ATSD-Signs, including the necessary image pre-processing (resizing, scaling) steps.
• Detection.ipynb shows how trained traffic sign detectors can be applied to on the ATSD-Scenes.
• Evaluation.ipynb illustrates how results from detection+classification pipelines can be evaluated w.r.t. ground truth annotations, in terms of per-class average precision, mean average precision (mAP), etc.
• weights contains trained weights of baseline classifiers and detectors.
• results contains pre-computed results of baseline detection+classification pipelines. See Evaluation.ipynb for information on how to make use of them.

The Austrian Highway Traffic Sign Data Set (ATDS) contains high-resolution traffic scene and -sign images acquired on Austrian highways in 2014 and annotated in 2020/2021. It consists of two main parts: ATSD-Scenes and ATSD-Signs.

ATSD-Scenes contains images of entire traffic scenes, where all traffic signs visible in a scene were manually annotated with a tight bounding box and information about the class of the sign and whether it possesses certain attributes (e.g., partly occluded, crossed-out, not normal to roadway, etc.). In total, there are 108 distinct traffic sign classes partitioned into 10 main categories, like mandatory, prohibitory, danger, etc.

Every image has a resolution of 1596x1196 pixels, and there are 7454 images in total. The images are split into the three disjoint sets Train, Test and Internal as follows:

Train and Test are part of the publicly available data, whereas Internal is kept private for being able to benchmark detection models in an objective way.

ATSD-Signs contains images of traffic signs, extracted from the full-scene images of ATSD-Scenes. Of the 108 traffic sign classes, the 60 most frequent appear in ATSD-Signs, as shown in the figure below.

There are 20,683 images in total, split into the three disjoint sets Train, Test and Internal. The splits correspond to the splits of ATSD-Scenes, i.e., a traffic sign image in Train was extracted from a scene in Train, etc. In the following table, Size refers to the length of the diagonal, in pixels, and is displayed as median (min-max).

Train and Test are part of the publicly available data, whereas Internal is kept private for being able to benchmark classification models in an objective way.

Performance on the public test set is measured by training on the training set and evaluating on the test set. Performance on the private internal set is measured by training on all public data (train+test) and evaluating on the internal set.

If you want to add the results of your models, please open a GitHub issue or submit a pull request. For evaluating your models on the internal set, please get in touch with us by opening a GitHub issue.

*Baseline refers to the model architectures presented in Classification.ipynb, with geometric+LED data augmentation as shown in Classification_Preparation.ipynb. Results are mean average precision (mAP), displayed as mean±SD over three independent runs.

Results in columns Categories refer to the detection+classification of categories 01 to 08. Results in columns Classes refer to the detection+classification of the 60 classes included in ATSD-Signs.

*Baseline refers to the model architecture presented in Classification.ipynb, with geometric+LED data augmentation as shown in Classification_Preparation.ipynb. Results are displayed as mean±SD over three independent runs.

• Alexander Maletzky, Nikolaus Hofer, Stefan Thumfart, Karin Bruckmüller, and Johannes Kasper. "Traffic Sign Detection and Classification on the Austrian Highway Traffic Sign Data Set". Data 8(1), 2023. DOI:10.3390/data8010016

  @article{ATSD_2023,
    title = {Traffic Sign Detection and Classification on the {A}ustrian {H}ighway {T}raffic {S}ign {D}ata {S}et},
    author = {Maletzky, Alexander and Hofer, Nikolaus and Thumfart, Stefan and Bruckm\"uller, Karin and Kasper, Johannes},
    journal = {Data},
    volume = {8},
    number = {1},
    year = {2023},
    issn = {2306-5729},
    url = {https://www.mdpi.com/2306-5729/8/1/16},
    doi = {10.3390/data8010016}
  }
  

If you have any inquiries, please open a GitHub issue.

This research was funded by FFG (Austrian Research Promotion Agency) under grant 879320 (SafeSign) and financed by research subsidies granted by the government of Upper Austria. RISC Software GmbH is Member of UAR (Upper Austrian Research) Innovation Network.