TWristAR - wristband activity recognition

TWristAR is a small three subject dataset recorded using an e4 wristband.   Each subject performed six scripted activities: upstairs/downstairs, walk/jog, and sit/stand.  Each activity except stairs was performed for one minute a total of three times alternating between the pairs.  Subjects 1 & 2 also completed a walking sequence of approximately 10 minutes.  The dataset contains motion (accelerometer) data, temperature, electrodermal activity, and heart rate data.   The .csv file associated with each datafile contains timing and labeling information and was built using the provided Excel files.

Each two activity session was recorded using a downward facing action camera.   This video was used to generate the labels and is provided to investigate any data anomalies, especially for the free-form long walk.  For privacy reasons only the sub1_stairs video contains audio.  

The Jupyter notebook processes the acceleration data and performs hold-one-subject out evaluation of a 1D-CNN.  Example results from a run performed on a google colab GPU instance (w/o GPU the training time increases to about 90 seconds per pass):

Hold-one-subject-out results

Train Sub
Test Sub
Accuracy
Training Time (HH:MM:SS)

[1,2]
[3]
0.757
00:00:12

[2,3]
[1]
0.849
00:00:14

[1,3]
[2]
0.800
00:00:11

This notebook can also be run in colab here.  This video describes the processing  https://mediaflo.txstate.edu/Watch/e4_data_processing.

We hope you find this a useful dataset with end-to-end code.   We have several papers in process and would appreciate your citation of the dataset if you use it in your work.