Abstract:
In the field of competitive boxing, performance is often evaluated by analyzing punch frequency and type, a process now enhanced by machine learning and Inertial Measurement Unit sensor data. Despite these advancements, punch range, a crucial factor influencing strategy and punch effectiveness, has been largely overlooked. To address this aspect, our work focuses on classifying punch types and ranges using various ML techniques trained on IoT sensors like IMU sensor data. We utilize spatiotemporal features, specifically power spectral density extracted from the four types of punch data, as input for the ML models. The models employed include Fine Decision Tree, Coarse Decision Tree, Linear Discriminant, Quadratic Discriminant, and Random Forest. In our comparative analysis of these ML models, we have found that the Random Forest classifier achieves the highest accuracy, accurately predicting punch types and their corresponding ranges with an impressive accuracy rate of 96.5%. This breakthrough in punch classification enables coaches and trainers to comprehensively assess a boxer's performance and design tailored training regimens for further improvement.