The DriveU Traffic Light Dataset (DTLD): Introduction and Comparison with Existing Datasets

This repository provides code for parsing the DriveU Traffic Light Dataset (DTLD), which is published in the course of our 2018 ICRA publication "The DriveU Traffic Light Dataset: Introduction and Comparison with Existing Datasets".

Paper

Paper see https://ieeexplore.ieee.org/document/8460737.

Download the dataset

The data can be downloaded from http://www.traffic-light-data.com/.

NEW v2 04/2021: json label format

.
├── DTLD                 # DTLD
    ├── Berlin           # Contains all Routes of Berlin
    ├── Bochum           # Contains all routes of Bochum
    ├── Bremen           # Contains all routes of Bremen
    ├── Dortmund         # Contains all routes of Dortmund
    ├── Duesseldorf      # Contains all routes of Duesseldorf
    ├── Essen            # Contains all routes of Essen
    ├── Frankfurt        # Contains all routes of Frankfurt
    ├── Fulda            # Contains all routes of Fulda
    ├── Hannover         # Contains all routes of Hannover
    ├── Kassel           # Contains all routes of Kassel
    ├── Koeln            # Contains all routes of Cologne
    ├── DTLD_labels_v1.0 # Old labels (v1.0) in yml-format
    ├── DTLD_labels_v2.0 # New labels (v2.0) in json-format
    ├── LICENSE          # License
    └── README.md        # Readme

Route structure

We separated each drive in one city into different routes

.
├── Berlin                # Berlin
    ├── Berlin1           # First route
    ├── Berlin2           # Second route
    ├── Berlin3           # Third route
    ├── ...

Sequence structure

We separated each route into several sequences. One sequence describes one unique intersection up to passing it. The foldername indicates date and time.

.
├── Berlin 1                    # Route Berlin1
    ├── 2015-04-17_10-50-05     # First intersection
    ├── 2015-04-17_10-50-41     # Second intersection
    ├── ...

Image structure

For each sequences, images and disparity images are available. Filename indicates time and date

.
├── 2015-04-17_10-50-05                                      # Route Berlin1
    ├── DE_BBBR667_2015-04-17_10-50-13-633939_k0.tiff        # First left camera image
    ├── DE_BBBR667_2015-04-17_10-50-13-633939_nativeV2.tiff  # First disparity image
    ├── DE_BBBR667_2015-04-17_10-50-14-299876_k0.tiff        # Second left camera image
    ├── DE_BBBR667_2015-04-17_10-50-14-299876_nativeV2       # Second disparity image
    ├── ...

Before starting

1. Check our documentation

Documentation is stored at /dtld_parsing/doc/. We give insights into the data and explain how to interpret it.

2. Change absolute paths

Do not forget to change the absolute paths of the images in all label files.

Loading the dataset

1. Download data & DTLD_Labels_v2.0.zip from https://cloudstore.uni-ulm.de/training/DTLD

2. In order to isolate dependencies between different Python projects and prevent conflicting requirements, it is a best practice to use virtual environments. More info can be found here. Create a virtual Python environment in which we can install dependencies required by the parsing scripts:

python3 -m venv .venv

3. After creation, activate the virtual environment like so:

source .venv/bin/activate

You can verify that the virtual environment is being used by running which python3, which shows the path of the python installation that you're currently using, if this shows a path ending in /dtld_parsing/.venv/bin/python3 the virtual environment is being used successfully

4. Run the setup script to install required dependencies using:

python3 setup.py install

5. Run the Python script to load the data, make sure that the label file & exported data are in the dtld_parsing folder, & provide the correct path to the label file <LABEL_FILE_PATH>:

python3 python/load_dtld.py --label_file <LABEL_FILE_PATH>

Analyze dataset

1. Download data

2. Create a virtual environment and install the dependencies as mentioned in steps 2, 3 & 4 in the previous guide. Skip if you've already done these steps as part of loading the data set.

3. Run the Python script to analyze the data, make sure that the label file & exported data are in the dtld_parsing folder, & provide the correct path to the label file <LABEL_FILE_PATH>:

python3 python/analyze_dtld.py --label_file <LABEL_FILE_PATH>

Citation

Do not forget to cite our work for the case you used DTLD

Citation:

@INPROCEEDINGS{8460737,
author={A. Fregin and J. Müller and U. Kreβel and K. Dietmayer},
booktitle={2018 IEEE International Conference on Robotics and Automation (ICRA)},
title={The DriveU Traffic Light Dataset: Introduction and Comparison with Existing Datasets},
year={2018},
volume={},
number={},
pages={3376-3383},
keywords={computer vision;image recognition;traffic engineering computing;DriveU traffic light dataset;traffic light recognition;autonomous driving;computer vision;University of Ulm Traffic Light Dataset;Daimler AG;Cameras;Urban areas;Benchmark testing;Lenses;Training;Visualization;Detectors},
doi={10.1109/ICRA.2018.8460737},
ISSN={2577-087X},
month={May},}