Detection of Proper Form on Upper Limb Strength Training Using Extremely Randomized Trees for Joint Positions

Learning how to perform a workout usually requires a little bit of research for people without experience, especially if it's in strength training or a workout that uses an equipment. Having an experienced instructor guide a trainee to perform such activity is enough, but with the advancement of computer vision, it opens up an opportunity to use technology to teach us without human intervention. In this study, Kinect V2 will be used to detect the joints of the subject and capture the starting and the end position of the workout. By tracking the position of the joints through the RGB-D sensor, the model will determine if the posture was performed correctly by the subject of these 5 (five) workouts: dumbbell lateral raise, dumbbell shoulder press, barbell front raise, dumbbell shrug, and barbell upright row. The data will be collected using the Java language and the KinectPV2 library to get the coordinates of the joints for each subject. Scikit-learn will be the chosen platform in the model creation for training and testing the datasets. Extremely Randomized Trees is a tree-based ensemble method for supervised classification and regression that will randomize. ERT will be used to the model proposed and compare it to the KNN.