Traffic Gesture Dataset
This page is the landing page for the traffic gesture dataset of the Institute of Microwave Engineering. The dataset consists of two main parts:
- Traffic Gesture Dataset: This is the main dataset with comprehensive radar recordings of eight traffic gestures. The radar recordings were accompanied by stereo camera captures, from which motion ground truth data has been extracted. This motion data is available, too.
- Out-of-Distribution (OoD) & Continuous Classification Dataset: This smaller dataset comprises continuous recordings of the eight traffic gestures as well as of other motion types. It is intended as validation and test set for continuous classification and OoD detection.
An overview over the two parts is given below. A more detailed documentation of the dataset can be found here (including all references to relevant publications). Moreover, a paper presenting approaches & baseline results for spectrogram-based classification, fast recognition, and OoD detection was submitted to EuRAD 2023 and is currently under review.
The data can be downloaded here.
For questions regarding the dataset, please contact nicolai.kern(at)uni-ulm.de.
Traffic Gesture Dataset
The traffic gesture dataset comprises radar data for 36 participants, recorded under various orientations. The eight gestures are visualized below. For each measurement, per-frame range-Doppler maps, per-frame target lists, and spectrograms are available in the gesture_measurements subdirectory. Each measurement is labeled with the gesture, participant ID, and orientation & position (as estimated by the stereo system). In addition, a ready-to-use dataset is available in the gesture_dataset folder. Further information can be found in the documentation.
The motion data are currently being processed and will be available at a later date.
Gesture 0 ("Fly")
Gesture 4 ("Push Away")
Gesture 1 ("Come Closer")
Gesture 5 ("Wave Through")
Gesture 2 ("Slow Down")
Gesture 6 ("Stop")
Gesture 3 ("Wave")
Gesture 7 ("Thank you")
OoD & Continuous Classification Datasets
Besides the traffic gesture recordings, the dataset comprises continuous measurements that contain either a) only gestures, b) only non-gestures, or c) gestures and non-gestures. An example of a measurement sequence with non-gestures is shown below. The cont_measurements folder contains recordings (RDMs, target lists, spectrograms) for six participants:
- backgrnd: Gestures and motions which are not part of the traffic gesture dataset (as in the video below).
- cont: Continuous gesturing; only gestures which are part of the traffic gesture dataset.
- contIdle: Continuous gesturing with breaks ("Idle").
- walk: No gesturing, walking around.
Per-frame labels are available as a file attribute within the files. The traffic gestures are assigned the labels 0-7; "Idle" is denoted as 10, gestures different from the traffic gestures ("Backgrnd") as 11, and "Walk" as 12. Out of these full measurements, ready-to-use datasets are extracted as described in the paper and the documentation. The datasets are in the folder cont_datasets, and can be used to test continuous classification and Out-of-Distribution detection.
Example of continuous gesture measurement
Example of gestures with "Idle"
Example of measurement with background motion