Human actions recognition on multimedia hardware using angle-based and coordinate-based features and multivariate continuous hidden Markov model classifier

Abstract

In this paper we have proposed human actions recognition schema that uses angle-based and state-of-the-art coordinates-based features and multivariate continuous hidden Markov model (HMM) classifier with Gaussian distribution. The main novelty of this paper besides presenting this approach is its evaluation on large (containing 770 actions) motion-capture dataset of various gym exercises. We have evaluated HMM with various number of hidden states and Gaussian mixture model classifier. We have performed PCA analysis of both features sets to justify our choices. We have also test 24 subsets of proposed angle-based features dataset in order to estimate which body joints are vital for correct actions recognition. The highest recognition rate in k-fold cross validation was 97 ± 14 % and was obtained for 4-states HMM with 10 angle-based features. The common problem with markerless motion capture vision - based systems is that if some parts of body surface are covered by another part it is impossible to perform accurate features measurements. Knowing this we have evaluated 24 subsets of proposed features dataset in order to estimate which body joints are vital for correct actions recognition.