ABSTRACT
Physical exercise is essential for healthy life since it has substantial physical and mental health benefits. For this purpose, wearable equipment and sensing devices have exploded in popularity recently for monitoring physical activity, whether for well-being, sports monitoring, or medical rehabilitation. In this paper, we proposed the sensor-based boxing movements detection and classification methods during shadow-boxing exercise which involves consecutive movements. A shadow-boxing exercise is a physical exercise in which you move your hands and feet as if you are boxing someone. The proposed method is evaluated on 10 participants aged between 21 to 25 years (10 males, 22.8 ± 1.4) and includes three martial art experienced and seven inexperienced persons. They are asked to follow instructions from an edited shadow boxing exercise video. We collected acceleration and angular velocity during the exercise sessions by attaching an IMU to the participant's right wrist. As an activity detection result, we achieved overall detection accuracy of 92.15%. The analysis inferred that using the right wrist as a sensor location makes detecting opposite hand punch movements difficult. As a result of a classification of 4 types of boxing movements (step-forward, step-backward, jab, cross), we achieved 63.07% of classification accuracy with Random Forest. The result showed the challenge of classifying shadow-boxing movements from a right-wrist-worn sensor. Although results showed that detecting and classifying a right-hand punch from a right wrist worn IMU is somewhat feasible by scoring 76%, other movements (a left-hand punch and footsteps) are not feasible with the proposed method. As far as we know, this is the first paper to propose sensor-based boxing movement detection and classification methods during shadow boxing exercises involving consecutive movements.
