This repository holds a small annotated dataset of motion capture sensor traces from subjects performing simple exercises (jogging, arm circles, toe touches, etc.).

Six sequences were collected from files available at mocap.cs.cmu.edu:

• Subject 13: trials 29, 30, and 31
• Subject 14: trials 6, 14, and 20

Each of the six sequences has been annotated to indicate which of a set of 12 possible exercises is being performed at each timestep.

The raw AMC data from these sequences was post-processed as follows:

• 12 sensor channels were kept as representative of gross motor behavior. Remaining channels were discarded.
• Each sensor channel was adjusted to have zero-mean.
• Each channel was block-averaged to a final frame rate of 10 fps (down from 120 fps in the raw data).

Matlab code to perform this post-processing is in preproc-code/.

Here's an example visualization of the data from subject 13, trial 30:

Each timestep consists of a 12-dimensional vector summarizing the positions and orientations of 12 body parts (see figure legend) during a 0.1 second interval.

The set of 12 possible exercise actions is listed below. Each letter A, B, C, ... in the figure indicates one of these behaviors.

    'JumpJack', 'Jog', 'Squat', 'KneeRaise', 'ArmCircle', ...
    'Twist', 'SideReach', 'Box', 'UpDown', 'ToeTouchOneHand', ...
    'SideBend', 'ToeTouchTwoHands' ...

Academic Citation

Fox, E., Hughes, M., Sudderth, E., and Jordan, M. "Joint modeling of multiple time series via the beta process with application to motion capture segmentation." Annals of Applied Statistics (2014). Volume 8, Number 3, Pages 1281-1313.

See Fig. 1 for a description of this dataset.

Dataset

Available as MAT file: mocap6.mat.

Description of fields stored in this file:

% * DataBySeq, 1D array of structs.
%    Each entry n is a struct for a single sequence, with fields
%    - X : 2D array, T x D
%        Each row t is observed data vector (size D) at time interval t
%    - Xprev : 2D array, T x D
%        Each row t is previous observation (size D) for time interval t
%        This field is used by auto-regressive models, ignored by others.
%    - TrueZ : 2D array, T x 1
%        Each entry t is the int label of human-annotated action at time t.
%    - filename : string
%        Indicates which record on mocap.cs.cmu.edu this came from.
%        Format: <subjectID trialID>, like 14_06 or 13_30.
% * ActionNames : 1D cell array of strings
%    each entry k is the name of the action referred to when TrueZ(t)==k
% * ChannelNames : 1D cell array of strings
%    each entry d is the name of the sensor for dimension d of X/Xprev

Credits

This repository was assembled by Mike Hughes (Brown University). Mike cleaned up the code and refined the annotations.

The dataset was originally assembled by E. Fox and E. Sudderth in 2009. E. Fox did the original annotation.

Thanks to CMU for making the raw data available.