Sports Motion Recognition Using MCMR Features Based on Interclass Symbolic Distance

Human motion and gesture recognition receive much concern in sports field, such as physical education and fitness for all. Although plenty of mature applications appear in sports training using photography, video camera, or professional sensing devices, they are either expensive or inconvenient to carry. MEMS devices would be a wise choice for students and ordinary body builders as they are portable and have many built-in sensors. In fact, recognition of hand gestures is discussed in many studies using inertial sensors based on similarity matching. However, this kind of solution is not accurate enough for human movement recognition and cost much time. In this paper, we discuss motion recognition in sports training using features extracted from distance estimation of different kinds of sensors. To deal with the multivariate motion sequence, we propose a solution that applies Max-Correlation and Min-Redundancy strategy to select features extracted with interclass distance similarity estimation. With this method, we are able to screen out proper features that can distinguish motions in different classes effectively. According to the results of experiment in real world application in dance practice, our solution is quite effective with fair accuracy and low time cost.