Prediction of Offensive Actions from Tankendo Video Images using Relative Frequencies of Image Features

The final goal of this study is to clarify, through system implementation, where and how an artificial system should look at human actions to be able to predict near-future results with high accuracy. The authors therefore collected video data from martial arts Tankendo practitioners and constructed an action prediction system for them. Its requirement specifications were to achieve 70% accuracy prediction within the first 0.25 seconds after movement onset, where a database was built using video-extracted features, specifically the bamboo sword tip height and area, and incorporating a calibration method to account for individual differences in middle guard stance and execution habits. Data from an expert practitioner was first used to compute a two-dimensional relative frequency distribution, and a Bayesian estimation and calibration approach was introduced, with 25 class interval and a cumulative probability threshold of 2.5. Calibration was performed by adjusting the database using initial sample and middle guard data from unknown subjects. As the results, the system was evaluated using data from the expert subject and two beginners. An accuracy of 75% was achieved with an output time of 0.3 seconds. Although calibration did not fully match expert performance in unknown subjects, but it consistently improved accuracy without delaying output.