Welcome. This dataset is free for non-commercial use. You can download content using the links below.

K3Da (Kinect 3D active) is a realistic clinically relevant human action dataset containing skeleton and depth data and associated participant information. With associated marker indicating noisy/outlier frames.

This dataset contains 54 participants, 26 young (<=59 years of age), 14 elderly (=>61 years of age) and 14 British Masters Athletes (=>61 years of age). Activities are based on the Short Physical Performance Battery (Guralnik et al., 2000) including: two leg jump, walking, sit to stand, and balance. In this dataset, all activities are recorded in accordance with the Short Physical Performance Battery protocols. All sequences have been recorded in a lab-based indoor environment with a single Kinect One 3D sensor, fixed to a tripod.

We seek to make this data available to the research community. We allow each file to download from our site. All we ask is that you cite the following paper when using the dataset:

D. Leightley, M. H. Yap, J. Coulson, Y. Barnouin and J. S. McPhee, "Benchmarking human motion analysis using kinect one: An open source dataset," 2015 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA), Hong Kong, 2015, pp. 1-7.

We welcome all comments and suggestions for future improvement for our dataset.

What file format do you use?

For each sequence we have recorded the skeleton stream at 30fps. We provide two files, a single matlab file consisting of all our skeletal data and the orginal raw skeleton files (text file). Each skeleton file contains the user position and pose in real world space. The skeletal file comprimes of 25 predefined joints as defined by Microsoft Kinect SDK, as well as the tracking state of both hands. The joint positions are given in X, Y, Z co-ordinates. Further, the file also includes the RGB mapped coordinates and tracking status. Finally, our Depth files are raw .bin files.

Noisy Data

We are aware that there is noise within our data. We have not excluded this from our dataset in the hope that it will be useful to the community. In the real world we have to handle noisy data, most notably within the health domain. If you find data that is very noisy and believe it should be excluded please do let us know. From our experiments we have found that the depth data is very noisy with partial occlusion, however we have provided it for visual reference.

Participant Characteristics

The characteristics of the dataset can be downloaded here. In this file we provide the age, gender, weight, height and an indicator of the type of participant (1=young, 2=British Masters' Athlete and 3=old).

How do I localise the co-ordinates?

The objective is to localise the skeleton in relation to the hip centre of the first frame. See my GitHub library for the codes.

May I have a copy of the entire dataset?

Yes. You may download our data as follows: K3Da Dataset. Please remember to cite our dataset if you use it in your work!

How have you recorded the data?

We conform to the SPPB protocol. All participants are asked to undertake the actions in their own time and in their own style.

What do the folder names mean?

Below is a list of actions contained within our dataset.

Warning Label/Code

A warning label/code is assigned when a participant encountered difficult in executing the trial. We provide this code as a guide as we are aware that no audio is provided therefore it may be difficult to understand the context of a failed attempt. It is up to the user of the dataset to decide if they want to include it when performing any analysis, evaluation or benchmarking.

Note

While we have taken GREAT care in recording this data. We do not guarantee the quality of the data. If you identify any problem data files please do inform us at: [email protected]