This dataset contains trajectories as well as body poses of pedestrians and cyclists in road traffic recorded in Aschaffenburg, Germany. It is appropriate for training and testing methods for trajectory forecasting and intention prediction of vulnerable road users (VRUs) based on the past trajectory and body poses. The dataset consists of more than 6526 trajectories of pedestrians and 1734 trajectories of cyclists recorded by a research vehicle of the University of Applied Sciences Aschaffenburg (Kooperative Automatisierte Verkehrssysteme) in urban traffic. The trajectories have been measured with the help of a stereo camera while compensating the vehicle's own motion. The body posture of the pedestrians and cyclists is available in the form of 2D and 3D poses. The 2D poses contain joint positions in an image coordinate system, while the 3D poses contain actual three-dimensional positions. A detailed description and evaluation of the pose estimation method can be found in [1]. In addition to the trajectories and the poses, manually created labels of the respective motion states are included. To read the provided data, unzip the file first. It contains one json file for each of the trajectories. Each json file contains the following data: vru_type: type of the VRU (pedestrian ('ped') or cyclist ('bike')) timestamps: UTC-Timestamps. The motions of the VRUs were recorded at a frequency of 25 Hz. set: Assignment to one of the three datasets train, validation or test. For pedestrians and cyclists, 60% of the data is used for training, 20% for validation and the remaining 20% for testing. During all splits, it was ensured that the distribution of the motion states is as similar as possible. pose2d: 2D poses with 18 joint positions in image coordinates with an additional uncertainty between 0 and 1 (third coordinate). Missing positions are encoded as 'nan'. pose3d: 3D poses with the trajectories of 14 joints in an three dimensional coordinate system. Missing positions are encoded as 'nan'. head_smoothed: Smoothed (by rts smoother) trajectory of the head in an three dimensional coordinate system. It is treated as ground truth position and must not be used as input for a prediction method. motion_primitives: One-hot encoded labels of the respective motion state. For pedestrians, a distinction is made between the states wait, start, move, and stop. For cyclists, the states wait, start, move, stop, turn left, and turn right are annotated. Python code for reading the data can be found on Github: github.com/CooperativeAutomatedTrafficSystemsLab/Aschaffenburg-Pose-Dataset Citation If you find this dataset useful, please cite this paper (and refer the data as Aschaffenburg Pose Dataset or APD): Kress, V. ; Zernetsch, S. ; Doll, K. ; Sick, B. : Pose Based Trajectory Forecast of Vulnerable Road Users Using Recurrent Neural Networks. In: Pattern Recognition. ICPR International Workshops and Challenges, Springer International Publishing, 2020, pp. 57-71 Similar Datasets Pedestrians and Cyclists in Road Traffic: Trajectories, 3D Poses and Semantic Maps Cyclist Actions: Optical Flow Sequences and Trajectories Cyclist Actions: Motion History Images and Trajectories More datasets Acknowledgment This work was supported by “Zentrum Digitalisierung.Bayern”. In addition, the work is backed by the project DeCoInt2 , supported by the German Research Foundation (DFG) within the priority program SPP 1835: “Kooperativ interagierende Automobile”, grant numbers DO 1186/1-2 and SI 674/11-2. References [1] Kress, V. ; Jung, J. ; Zernetsch, S. ; Doll, K. ; Sick, B. : Human Pose Estimation in Real Traffic Scenes. In: IEEE Symposium Series on Computational Intelligence (SSCI), 2018, pp. 518–523, doi: 10.1109/SSCI.2018.8628660