An Trial of Predicting Effective Strikes in Martial Arts using Image Features based on Shifting Patterns of RGB Intensity

Our goal is to understand, through system implementation, where and how it 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 practitioners and constructed an action prediction system for them. In this study, the authors proposed a new system for high-accuracy and early prediction of martial arts tankendo techniques. This system achieves motion prediction by dividing input images into 64×64 pixel blocks and analyzing the temporal change patterns of RGB values across three consecutive frames of image blocks. We enhanced prediction accuracy by encoding the position information of image blocks using trigonometric functions and fusing it with RGB features to preserve spatial information. Additionally, to accommodate subjects with different physical builds, we introduced an adaptive calibration method that maps the distribution patterns of active blocks to a standard model. This system is designed to utilize finely divided RGB pattern features and perform predictions through deep learning. However, experimental results revealed that the system failed to achieve this initial goal. The earliest prediction times were 0.462 seconds before completion for head strike, 0.363 seconds for throat thrust, and 0.330 seconds for torso thrust. What the series of results showed was that the proposed system only discriminated between the three types of techniques in the appearance, where the final outcome was generally known, and could not be said to have predicted the final outcome of the technique in the appearance, where the final outcome was still not well known.