Bicep Detection Using Electromyography and 3-Axis Accelerometer Data

The Research Paper "Detection of Bicep Form Using Myoware and Machine Learning" based on the novel dataset has been recently accepted in September 2022 and is being published in SCOPUS Indexed SPRINGER Book Series “Lecture Notes in Networks and Systems”

Abstract of the Research paper:

Many people have been exercising at home since the beginning of the COVID-19 period. Due to this, they haven’t had the supervision of a professional trainer who could correct them, rectify their mistakes, and prevent from harmful injuries. One very common workout that people frequently do is the bicep curl exercise, yet they fail to maintain the right posture without realizing it and end up straining their back muscles and pulling their shoulders too far forward. This is very dangerous as it could result in long-term back pain and tendonitis. Considering this issue, a methodology has been proposed for the people at home to continue their regular exercises without any professional gym trainer and gym environment. This research is oriented toward the correction of bicep form during the bicep curl exercises and preventing injuries. The acquisition fragment is designed with Myoware, an open-source electromyography sensor along with a 3- axis accelerometer for capturing the essential segments required for the analysis. Naïve Bayes, Logistic Regression, K-Nearest Neighbor, Decision Tree, and Random Forest Classification models have been used to perform the classification of the data acquired. The analysis of the novel dataset attained has been effective for determining the model with the best results to detect the bicep form. The Random Forest Classifier has yielded the highest individual accuracy of 90.90%. Ultimately, an app prototype has been developed using the MIT app inventor platform. It has been integrated with a Bluetooth module and google firebase for showcasing and collecting data in real-time from the sensory modules to the person working out and simultaneously storing them in the database for the deployment of this test set in the Machine Learning Model. With this, the results of the accuracy and performance of a person are then acquired on the app for expressing the nature of their workout session.

About the Dataset:

A dataset was generated for this research which consisted of four independent variables and a dependent variable (Total of five columns). The first column represented the data acquired from the EMG signals via the Myoware muscle sensor. The signals in the dataset captured different types of contractions of the muscle during exercise. When the person exercised in the wrong form, the bicep was stretched abnormally, and the value generated by the signals differed from the values generated by the correct form of exercising. However, anticipating the right form of the bicep based only on the EMG signals was unreliable as every person has a different muscle strength mass. While recording the exercise, some value that was obtained from a bodybuilder’s arm during an aberrant movement of the bicep muscle was conflicting with the value acquired from a skinny person’s arm during the best form and contraction of their bicep muscle. In order to tackle this anomaly, three more independent variables were introduced that dealt with the arm's position and orientation. The 1st,2nd, and 3rd columns contained the data acquired from the X, Y, and Z axis of the accelerometer. The Dataset Image shows that The Myoware consisted of integer values while the data type of the values acquired from 3 axes were floating points and the output was an object, representing a sequence of characters. The accelerometer was placed at a 90-degree angle on the deltoid muscle. As the exercise began, the sensor started to capture the movement and orientation of the arm. Variant values from the accelerometer were collected during the right and wrong forms of the bicep curl exercise. For creating this dataset, 10 volunteers were selected who were instructed to perform the exercise. The volunteers included people having a good build and regular gym-goers, people having skinny built and who rarely ever exercised, athletes and sports players, and healthy people who aspire to start going to the gym soon. They used dumbbells from 2 kg up to 15 kg with which a diverse and adequate dataset was gathered. A Data Logger script was coded using the Python Serial Module with an Arduino nano connected through the USB. Data consisting of 5000 variant values were collected from the Myoware Sensor and Accelerometer simultaneously using the Python script. Consequently, it was added to the CSV file for using it with Excel and other data analytic tools.