UC-3D Motion Database
Your open-source multisensor 3-D database for action recognition.
About
Action analysis and recognition is a very active research topic. The research group of the Mobile Robotics Laboratory from the Institute of Systems and Robotics, University of Coimbra releases the University of Coimbra 3-D Motion Database (UC-3D) so to promote this research. The database was developed in the context of an Action-based Person Recognition research work, for which it is also found suitable for research on action-based biometrics.
Available data types encompass high resolution Motion Capture, acquired with MVN Suit from Xsens and Microsoft Kinect RGB and depth images. This diversity intents to provide the community with an homogeneous dataset for fair comparison between methods using different datatypes.
In the UC-3D Motion Database researchers will currently find 11 different activities, performed by 13 different actors. 6 of these actions are interactive, containing 2 different persons, for which the MVN data is only recorded for one of them. Each person performs 3 trials for each different action. Typical action duration is 3 seconds.
Body skeleton drag reconstructed from
MVN Cartesian Data
RGB frame example
Depth frame example
Acknowledgements: Special thanks the following persons, who have contributed as actors for data acquisition:
Structure
The database is composed of 3 different data types, stored using ZIP compression, where the format of a file name is 'ddd-aaa-ppp.zip' and:
ddd - data type which can be:
RGB – RGB images
DEP – DEP images
-
MVN – XML file containing information for the 17 Inertial Measuring Units of the MVN suit
With respect to the XML with the MVN data, it contains data as in the following example:
<mvnxInfo frameCount="419" frameRate="120" segmentCount="23" recDate="1375790676.105" originalFilename="MVN-BND-p01.mvn" velData="1" accData="1" angVelData="1" angAccData="1" />
aaa - performed action:
BND – bend
CNV – conversation (2 person interaction)
HSH – handshake (2 person interaction)
HUG – hugging (2 person interaction)
JMP – Jumping
PCH – Punching (2 person interaction)
PCD – punched (2 person interaction)
PSH – Pushing (2 person interaction)
RUN – Running
WLK – Walking
SIT – Sitting/Standing Cycle
ppp - enumerated from p01 to p13:
identifies main actor, whose MVN is captured for.
Download
Remark 1: data for p10 to p13 will be uploaded soon.
Remark 2: RGB data for the conversation action will be uploaded soon, due to filesize dimensions (+1.5GB).
Remark 1: These actions were acquired under the scope of the Prometheus project.
- Scenario Type:
- Outdoor
- Description:
- Normal behaviour, taking money at ATM (2 persons in a row, change of person)
- Devices Used:
- Fire Wire Camera
- Microphone Array
- Download:
- ATMD 01
- Scenario Type:
- Outdoor
- Description:
- Theft: Person 1 takes money at ATM, a person 2 grabs it out of the hand, runs away
- Devices Used:
- Fire Wire Camera
- Microphone Array
- Download:
- ATMD 02
- Scenario Type:
- Outdoor
- Description:
- Theft: similar action as in 2
- Devices Used:
- Fire Wire Camera
- Microphone Array
- Download:
- ATMD 03
- Scenario Type:
- Outdoor
- Description:
- Normal behaviour, taking money at ATM
- Devices Used:
- Fire Wire Camera
- Microphone Array
- Download:
- ATMD 04
Licence
Please note that the database is publicly available for non-commercial use. If you use this database in your publications at the moment, please use the following bibtex code:
@article{SANTOS2015568,
title = "Trajectory-based human action segmentation",
journal = "Pattern Recognition",
volume = "48",
number = "2",
pages = "568 - 579",
year = "2015",
issn = "0031-3203",
doi = "https://doi.org/10.1016/j.patcog.2014.08.015",
author = "Luís Santos and Kamrad Khoshhal and Jorge Dias
}
Support Software
We provide a MatLAB Toolbox, which allows extracting data from the MVNX files into a structure. You can find it here.
Contact
If you have any questions regarding the database or how to work with toolbox, please email: luis@isr.uc.pt