Myo dataset made for the reaseach presented in my paper about "Person-Independent sEMG Gesture Recognition Using LSTM Networks for Human-Computer Interaction".
This repository contains sEMG Data of 13 subjects recorded with the Myo Armband. Subjects were instructed to wear the armband on their dominant arm, like this:
The gestures that were recorded, are the ones from the Rock, Paper, Scissors game:
Each recording was given it's own folder.
The folder are named like this: <subjectId>_<arm>_<index> (also called the label)
Where arm r stands for the right arm.
Each kind of data is stored in it's own file per recording.
| File Ending | Description |
|---|---|
<label>-<gestureType>-<index>-emg.csv |
Raw EMG Data from the Myo armband of the eight sensors |
<label>-<gestureType>-<index>-pose.csv |
Pose data (detected poses by the built-in Myo detecion) |
<label>-<gestureType>-<index>-orientation.csv |
IMU Orientation data |
<label>-<gestureType>-<index>-meta.csv |
Meta Data (see below) |
Each gesture recording comes with a meta data description file.
Example with explainatory comments:
label: s6_r_1 # same as the folder
gesture: scissors # which gesture the user was told to record
index: 8 # 8th gesture of type scissors from this session
date: 21/08/2020
time: 05:49:30
arm: 0 # arm 0 = right, arm 1 = left
arm_direction: 1 # -1 if the armband was worn in reverse
rssi: -53 # signal strength
battery_level: 97 # 100 means full battery
mirror_left_arm: True # wether the recording applicatin automatically mirror the data if the user was left handed
files: # which data was recorded; e.g. pose data might miss, if no pose was detected
- s6_r_1-scissors-8-pose.csv
- s6_r_1-scissors-8-orientation.csv
- s6_r_1-scissors-8-emg.csv
In order to generate labeled data, a Python application interfacing with the Myo was implemented. The data is recorded while showing an icon and text label describing the gesture. Each gesture is recorded for two seconds following a 0.5 second break where subjects are supposed to relax their arm. The two-second time-span was chosen because several tests showed that it is the amount of time required for most persons to perform the transition from rest pose to the wanted gesture and back. In one recording iteration, each gesture is recorded ten times, which makes a total of 30 gestures per iteration. Therefore, one iteration takes 75 seconds. Those values were chosen after initial testing, which showed that subjects quickly get exhausted performing the gestures. After a short break, without taking the armband off, a second iteration was performed in the same session. This is then repeated once more. Some participants were recorded a second time: After a long break, the armband was put on differently, and the process described above was repeated. In total, each participant contributed three (one session) or six iterations (two sessions). One additional participant recorded 30 datasets in 10 sessions on different days to create a highly specialized dataset.
Thirteen healthy subjects aged between 18 and 58 years participated in the data acquisition process. Four of them were female, and nine candidates male. One participant was left-handed. All subjects agreed to the data recording and analysis by writing. Before starting the recording session, participants filled out the consent form, which states what is being recorded and how the data is used. The Myo armband was then put onto the thickest part of the forearm, just below the elbow on the dominant arm. Then the subject waited until the connection icon on the armband stopped blinking, which indicated that the armband warmed up. After performing the Myo sync gesture, which is needed to detect the arm wearing the armband, the Myo is ready to use. Subjects were told to start from a rest pose, performing the gesture and reaching the highest intensity when the progress bar indicates that half the time passed.
There are multiple reasons for always using the dominant arm during the recording. First, a possible user interface using one Myo armband would be controlled with the dominant arm because it feels more natural to the user. But also not using the dominant arm could introduce irregularities into the dataset because users are not used to performing similar gestures with their non-dominant arm. Also, differences in the data resulting from recording the left and right arm are non-existent since the current arm is detected, and the data mirrored accordingly.
All data provided by the Myo armband is stored in multiple files labeled with a unique but randomized identifier. Each file contains roughly 400 ± 5 entries since the recording took two seconds, and the data was sampled with. Each entry contains an index, a timestamp, and one value for each of the eight sensors. and other data along with meta-information (label, recording time, left or right arm, battery level, and connection quality) is stored in other files.