Karate Kata Style Classification Using Pose Landmarks and Deep Learning
Mahmoud Daker, Farida Elsayaad, Ayman Atia · 2023
This paper aims to shed light on the efficiency of using LSTM in the binary classification of Karate Kata players’ styles. In addition, this paper showcases the results of the proposed LSTM model on the MPII Human Pose Dataset. Since our focus in this research is LSTMs, our paper also showcases the different approaches we applied in order to reach the best classification results. Moreover, the approach we settled on was to classify each video as a whole, using a player’s landmarks as input for the LSTM model. This study concludes that in comparison to segmenting the video and classifying the moves individually, the approach of classifying the video as a whole showed significant improvement. It was also concluded that using LSTMs was efficient in the task of classifying the Karate Kata style of the player with an accuracy of 90.3%. Furthermore, the model’s precision, recall, and F1-Score were 94%, 90%, and 92% respectively. It is also worth mentioning that different approaches such as using classical machine learning models also showed promising results. Further research in the field is needed in order to reach a more concise result.