Support learning vovinam exercises based on computer vision

Abstract

Computer vision has many applications which has attracted many researchers, especially with the problems of recognizing actions, postures, movements. This Thesis offers a method to support students to perform correct postures during martial arts practice. We have collected and labeled data about the movements of a traditional Vietnamese martial art called Vovinam. In the original paper of ST-GCN, before input into the model, we need to transform videos to the sequence of keypoint positions by frame to be handled in the next phase; it seems to be that normally the transform phase didn't reach effective performance. Therefore, the purpose of this thesis is to improve the ST-GCN model in terms of input. Firstly, we use a sequence of recently released techniques to extract the skeleton and its key points from the input video. Then, the sequence of keypoint positions by frame will be inputted into the deep learning architecture based on the ST-GCN model and the output will be the determined action. On our dataset, adding the input processing stage to the recognition model has yielded much better results than applying the original model. The final accuracy is 99.23%, showing that the model has the potential to be applied in practice.