Nihon Kikai Gakkai ronbunshu (Nov 2021)
Deep learning method for karate motion identification using inertial sensor data
Abstract
Small inertial sensors having low cost and simple structure have recently been developed owing to improvements in microelectromechanical systems. Therefore, inertial sensors are being widely used in various sports for quantitative measurements and athletes’ evaluation. However, no wearable device has been specifically designed to capture karate movements such as punching, blocking, and kicking, and these movements are being recorded manually in the field of karate training. Automated recording and identification of karate movements using wearable sensors can lead to improved training and performance, therefore we developed an identification method for various karate movements using inertial sensor data. First, we created a dataset of karate movements by acquiring measurements from 22 karate players at the karate club of Waseda University. Their competitive levels were wide-ranging, from Japanese local competition level to international competition level. Inertial sensors were attached to five parts of each participant’s body (right hand, left hand, right foot, right leg, and waist), and basic karate movements (reverse punch, upper block, and front kick) were performed by each participant for measurement. Then, we established the dataset containing inertial sensor data paired with the correct karate movement label. Next, we designed a deep learning network based on long short-term memory and trained the network using part of the dataset. We evaluated the network using leave-one-out cross-validation, obtaining the F1-scores of 0.90 for right reverse punch, 0.89 for left reverse punch, 0.88 for right upper block, 0.86 for left upper block, 0.90 for right front kick, and 0.90 for left front kick. In addition, the micro-average of the F1-score was 0.89. Therefore, high identification accuracies were achieved for all types of karate movements.
Keywords
