1College of Intelligence and Computing, Tianjin University, Tianjin 300350, China
2Institute of Disaster and Emergency Medicine, Tianjin University, Tianjin 300072, China
* Corresponding author
Emergency Medical Services (EMS) play a critical role in acute emergencies, yet their effectiveness is often limited by professional complexities. For example, a European study on out-of-hospital cardiac arrest (OHCA) found survival rates below 10%, primarily due to delayed responses and insufficient bystander intervention. Existing datasets for medical movement analysis have largely focused on basic patient actions like lying and standing. The NTU dataset includes 2D joint data for actions such as sneezing and covering the head in everyday sickness scenarios. In daily life scenarios uses motion recognition technology to monitor patients’ postures, and in hospital environments uses LiDAR for human posture and motion recognition. However, there is a significant gap in research on the actions of rescuers in medical emergency procedures. Creating a 3D dataset of medical emergency procedures can provide data support for the analysis or generation of emergency medical procedures, thereby facilitating the dissemination of emergency medical practices.
The data collection process for EMP3D is meticulously designed to accurately capture the movements of medical personnel during emergency procedures. Six GoPro cameras are strategically positioned around a simulated medical environment to capture multiple critical perspectives. The video streams from these cameras are synchronized using sound signals, ensuring precise alignment across all six viewpoints. Initially, the 2D posesof six synchronized viewpoints are estimated based on rtmpose and dwpose. Subsequently, joint points across these viewpoints are matched through 4D association to derive 3D joint coordinates. We have created theTracking module to track the positions of the rescuers and patients in each frame. The "tracking" module primarily involves operations such as trajectory initialization, feature vector construction, cost matrix computation, and linear matching. Ultimately, an accurate 3D SMPL-H parameterized model is obtained via a two-stage optimization method.Manual inspections are conducted for both 2D and 3D keypoint results of each frame. Any identifiedissues are manually adjusted to refine the results. Additionally, the 3D SMPL-H results for each frame are also manually inspected. Although the manual inspection process is time-consuming, it ensures a high-quality dataset.
This dataset is intended to support academic research and non-commercial applications. Please read and agree to the following terms before downloading and using this dataset.
1.Scope of License: This dataset is provided exclusively for academic research, teaching, and personal learning purposes. Any commercial use or profit-making activities involving this dataset are strictly prohibited without explicit written permission from the copyright holder.
2.Copyright and Intellectual Property: This dataset is protected by relevant laws and regulations. We retain all rights, ownership, and interests in all contents of the dataset. Except for the granted use mentioned above, this agreement does not grant you any explicit or implicit rights to patents, trademarks, or other intellectual property.
3.Prohibited Actions: Downloaders are prohibited from selling, reselling, leasing, renting, distributing, or commercially exploiting this dataset in any form. Additionally, modifying, adapting, translating, reverse-engineering, decompiling, or creating derivative works from the dataset is not allowed. Downloaders are also prohibited from attempting to access the source code (if applicable) by any means.
4.Citation Requirement: If you use this dataset in your research, we request that you appropriately cite our work and provide a link to the original location where the dataset was released when publishing your results.
Please carefully consider the above terms before downloading. If you do not agree with these terms, please do not download or use this dataset. If you agree, you may proceed to download the dataset by visiting https://zenodo.org/records/14548723.
The EMP3D dataset is structured as follows:
EMP3D ├── CPR │ ├── yyyymmdd_xxxxxx │ │ | │ │ ├── keypoints │ │ ├── smpl_parameters │ │ ├── cameras │ │ ├── videos │ │ └── synchronization_information │ └── ... ├── Hemostasis │ └── ... ├── Bandaging │ └── ... ├── Heimlich │ └── ... └── Fracture_fixation └── ...
If you have any questions or suggestions, please feel free to contact us at: lik@tju.edu.cn. Address: Building 55, Tianjin University Beiyangyuan Campus, 135 Yaguan Road, Haihe Education Park, Jinnan District, Tianjin