Github Repository
Click here!
Welcome!
Welcome to the homepage of the DriveU Traffic Light Dataset (DTLD). We present a dataset which addresses to researchers in the field of traffic light recognition/detection. The main contributions of this dataset are:
- Annotation quantity: We provide the highest number of annotated traffic lights compared to other existing datasets (11/2018).
- Annotation quality: We provide accurate and consistently annotated objects due to using an appropriate labeling tool. Furthermore all annotations have novel attributes such as viewpoint orientation, pictogram and relevancy information. Track identities are assigned to group unique traffic light instances.
- Additional sensor data: In addition to camera images we provide stereo camera images, calibration data and vehicle data such as GPS, velocity and yaw rate.
Dataset Overview
DTLD contains more than 230 000 annotated traffic lights in camera images with a resolution of 2 megapixels. The dataset was recorded in 11 cities in Germany with a frequency of 15 Hz. Due to additional annotation attributes such as the traffic light pictogram, orientation or relevancy 344 unique classes exist. In addition to camera images and labels we provide stereo information in form of disparity images allowing stereo-based detection and depth-dependent evaluations.
Labeling down to the resolution limit
Our sophisticated labeling tool allows to annotate traffic lights down to a width of 5 pixels or below. Therefore DTLD consists of a high number of very small objects (Figure 2). This makes the dataset also predestined for researchers working on small object detection.
High aspect ratio variance
By also annotating traffic lights consisting of one, two (e.g. pedestrian traffic lights) or four light units (e.g. bus/tram traffic lights) DTLD has a higher aspect ratio variance than other datasets.
High regional variance
One important property of DTLD is its high regional variance (Figure 4), i.e. the distribution of the labels in the 2D image. This was reached by only adding non-static scenes to the dataset. Only static scenes, in which at least one traffic light state changes are added.
Labels
We provide annotations in terms of bounding box coordinates (top left corner, width and height) but also extensive traffic lights properties, namely:
- Viewpoint orientation
- Relevancy
- Installation orientation
- Number of light units
- State
- Pictogram
These properties are expressed by a six digit class identity in v1.0 and as an attribute dictionary in DTLD v2.0. In addition each label has a track identity (useful for tracking evaluations). For each image, a timestamp and vehicle data (GPS, velocity and yaw rate) is available.
Class identity
Digit I: Viewpoint orientation
Viewpoint orientation describes the orientation of a traffic light with respect to the ego-vehicle. There exist 4 tags:
- Front: valid for the ego-vehicle and the state/pictogram is mostly visible
- Back: traffic lights are valid for the oncoming traffic
- Left: traffic lights turned to the left, mostly pedestrian traffic lights
- Right: Traffic lights turned to the right, mostly pedestrian traffic lights
Digit II: Relevancy/Occlusion
- Relevancy: A traffic light is relevant if it is valid for the planned route of the vehicle. Traffic lights of the next intersection are not tagged as relevant.
- Occlusion: Concludes occluded and truncated traffic lights
Digit III: Installation orientation
- horizontal: Horizontally orientated traffic lights, mostly in Asia or US
- vertical: Vertically orientated traffic lights
Digit IV: Number of light units
Self-explanatory
Digit V: State
Self-explanatory. Red-Yellow is a transition state in Germany indicating the change from red to green.
Digit VI: Pictogram
Expressing the mask of the lamp (circular, arrow + direction, pedestrian, bike ...)
Track identity
Each unique traffic light instance has an assigned track identity during one sequence. One sequence consists of one drive to an intersection until passing it.
Vehicle data
DTLD provides temporal information in the form of one timestamp per image as well as local information in the form of GPS information.
How to use the dataset
We provide tools and scripts for using the dataset in C++, Python and MATLAB. A documentation is added as well.
Download
Information
Click here for registration and download.
Citation
For the case you used DTLD for your scientific work please do not forget to cite our ICRA2018 paper.
Publications based on DTLD
2016:
Fregin et al.: A closer look an traffic light detection evaluation metrics, ITSC 2016
2017:
Fregin et al.: Three ways using stereo vision for traffic light recognition, IV 2017
Fregin et al.: Feature detectors for traffic light recognition, ITSC 2017
2018:
Müller et al.: Detecting Traffic Lights by Single Shot Detection, arXiv preprint 2018
Please contact us to add your publication based on DTLD!
Login page for Download
Feedback/Fragen/Probleme jeglicher Art
Feedback? Questions? Any problems/errors? Do not hesitate to contact us!
Reporting Labeling Errors
We are currently still working on improving our dataset. You can help us by reporting not labeled traffic lights, inaccurate labels, or wrong class identities. We would appreciate a json file with the following structure:
[
{"path": "./Berlin/Berlin1/2015-04-17_10-50-05/DE_BBBR667_2015-04-17_10-50-13-633939_k0.tiff", "objects": [{"y": 313, "x": 1353, "height": 19, "comment": ["error", "object_error"], "width": 6}]},
{"path": "./Berlin/Berlin1/2015-04-17_10-50-41/DE_BBBR667_2015-04-17_10-50-46-968138_k0.tiff", "objects": [{"y": 285, "x": 1306, "height": 25, "comment": ["error", "classid_error"], "width": 7}]}
]