A Study on Human Motion Assessment Method by RGB-D Video Processing Based on Support Vector Regression

Abstract

In recent years, the demand for the system that evaluates human motion is increasing. Studies of recognizing motion are actively being conducted, but not many works on quantitative assessment of human motion can be seen. This paper proposes an automatic algorithm that assesses human action for multiple scoring items by learning evaluation scores for each item. More specifically, to obtain the features, the proposed algorithm performs DCT (Discrete Cosine Transform) to the sequences of the 3D coordinates of human joints. The obtained features are learnt by Support Vector Regression. As a result of the leave-one-out method based experiments that give 10 points as the full score to each action, a conventional method and the proposed method achieves the mean error of 0.59 and 0.26, respectively. It turns out that the proposed method can improve the evaluation accuracy.