Automated Dance Scoring Algorithm Using Alignment and Least Square Approximation with Fractional Power of Joint Features

Automated motion evaluation has become popular in exercise training and entertainment. In this study, an advanced automatic dance scoring algorithm is proposed. First, to avoid misjudgment from misalignment, space and time alignment are assessed. Then, instead of using the whole video frames as the input, we apply the joint information, including the relative locations, the moving velocities, the orientations, and the areas between the joint lines. To make the features more flexible and magnify the detail difference, we take the fractional powers on input features. The correlation coefficients are calculated for feature selection, and a nonlinear analysis is introduced to determine the angle difference. The least mean square error approximation is also applied to determine the linear combination coefficients of the features. The difference between the ground truth and the interpolated results from the regression line is minimized using the input features. The proposed algorithm accurately predicts dancing scores.