Development of the evaluation algorithms for karate skills using IMU sensors

Abstract

Our research aimed to develop algorithms to evaluate the quality of karate motions. Kata is the representation of karate’s self-defense techniques strung together into a performance routine. Kata is judged based on several technical and physical criteria including speed, strength, focus, breathing, balance, and rhythm. For this reason, evaluation of karate motions is challenging. In this research, we created a novel dataset of referee scores and inertial sensor data of karate movements. The subjects were 22 members (15 males, seven females, age 20 ±1.3 years) of Waseda University’s Karate club who competed at the international and regional level. Inertial sensors were attached to five body parts (forearms, lower legs, and waist) and the subjects performed fundamental movements in karate (reverse punch, upper level block, and front kick) as the target actions. Subjects performed 30 trials for each action. The quality of each action was scored by an official referee as the ground truth. Also, the quality of each action was scored by subject's self-assessment as the comparison. The resulting data was distributed into the learning dataset and the evaluation dataset. Next, we developed a classifier that evaluates the quality of each action in the learning dataset in three stages. First, the importance of each feature was judged using ensemble learning. The classifier then evaluated the karate motions using handcrafted features of high importance. Finally, the classifier constructed strong classifiers by combining weak classifiers. As a result, our evaluation method was applied to the test dataset. The matching rate of the estimated value and the ground truth was 0.830 ± 0.067 (Mean ± SD). Also, the accuracy of self-assessment was 0.504 ± 0.039. Also, there was significant difference at the 1%.

Keywords

• machine learning
• random forest
• imu sensor
• karate
• quality evaluation