Classification of Karate Kicks with Hidden Markov Models Classifier and Angle-Based Features

Human actions classification is important task of modern vision-based and sensor-based motion capture (MoCap) systems. This research is part of our project on human action recognition and analysis aiming especially at sport and rehabilitation activities. We proposed a HMM classifier with angle-based features to recognize various types of karate kicks. We examine various number of hidden states and number of dimensions reduced by Principal Component Analysis (PCA). In contrary to the previously published researches we want to design a classifier capable to recognize very similar (in the sense of spatial trajectories) human actions recorded by precise motion capture hardware. The best classification results with total recognition rate 0.92 was obtained for HMM with 7 hidden states (without dimensionality reduction), and for 6 hidden states after reducing number of dimensions with PCA to 5. Both the complete dataset and source code we used in this research can be downloaded to reproduce the research results.