Pediction of attacking motions from video images of martial art Tankendo using support vector machine

This study aims to develop a Tankendo offensive actions prediction system that can achieve the same performance level as known offensive actions when an unknown attacker performs a strike. The data were collected from a highly skilled Tankendo practitioner, including 100 strikes to the head, 100 to the throat, and 100 to the torso. The dataset was constructed by extracting data from the one-second period before each strike was completed, with 240 strikes used for system development and 60 for parameter optimization. During video processing, the system tracked the attacker’s collar, the tip of the bamboo sword, and the guard to compute features such as angle, distance, velocity, and acceleration. Principal Component Analysis was then applied to reduce the feature dimensions to nine. Additionally, to handle sections that were difficult to classify, an “Other” class was defined, and its labeling position was set as a tunable parameter. The number of consecutive frames used before the final output was also set as an optimizable parameter and was evaluated through a full parameter combination analysis using 60 strike sequences. To ensure predictions were made before the mid-phase of the motion, the labeling position was ultimately set to 0.5 seconds before strike completion, with three consecutive frames used as the final decision criterion. When analyzing two unknown attackers, we first measured their strike durations and selected the optimal parameters based on the known data. The evaluation included two attackers with longer strike durations than the known attacker. The experimental results indicate that for Subject A, the system’s average output time was −0.4 ± 0.1575 seconds before strike completion, with correct outputs at −0.3715 ± 0.1328 seconds and an accuracy of 50 percent. For Subject B, the system’s average output time was −0.4829 ± 0.1123 seconds before strike completion, with correct outputs at −0.4983 ± 0.1127 seconds and an accuracy of 66.67 percent. Therefore, at this stage, the system has not fully achieved the same performance level for unknown attackers as it has for known attackers.