Application of assistive computer vision methods to Oyama karate techniques recognition

Abstract

In this paper we propose a novel algorithm that enables online actions segmentation and classification. The algorithm enables segmentation from an incoming motion capture (MoCap) data stream, sport (or karate) movement sequences that are later processed by classification algorithm. The segmentation is based on Gesture Description Language classifier that is trained with an unsupervised learning algorithm. The classification is performed by continuous density forward-only hidden Markov models (HMM) classifier. Our methodology was evaluated on a unique dataset consisting of MoCap recordings of six Oyama karate martial artists including multiple champion of Kumite Knockdown Oyama karate. The dataset consists of 10 classes of actions and included dynamic actions of stands, kicks and blocking techniques. Total number of samples was 1236. We have examined several HMM classifiers with various number of hidden states and also Gaussian mixture model (GMM) classifier to empirically find the best setup of the proposed method in our dataset. We have used leave-one-out cross validation. The recognition rate of our methodology differs between karate techniques and is in the range of 81% ± 15% even to 100%. Our method is not limited for this class of actions but can be easily adapted to any other MoCap-based actions. The description of our approach and its evaluation are the main contributions of this paper. The results presented in this paper are effects of pioneering research on online karate action classification.