Trained table based recognition & classification (TTRC) approach in human motion reconstruction & analysis.

In this paper, first we discuss human motion analysis using the temporal template methodology. This methodology deals with the creation of motion history images (MHIs). Hu moment invariants are calculated from MHIs, for feature description. Two types of training datasets based on Hu moment invariants, have been developed. One training dataset is of 105times7 elements and other consists of 200times7 elements. Secondly a new simple approach for motion recognition and classification called trained table based recognition & classification (TTRC) has been proposed. In TTRC approach, instead of using the training datasets, a simple data table has been trained. The training process comprises of the behavioral study of seven phi values of Hu moment invariants. We make performance evaluation of TTRC with other classification techniques using these training datasets in the context of accuracy, success rate, time & speed and memory capacity. The classifiers used in this paper are K-nearest neighbor (KNN) and fuzzy K-nearest neighbor (FKNN) classifiers with values of K = 1,3,5, Mahalanobis distance (MD) classifier, linear Bayes Gaussian (LBG) classifier, quadratic Bayes Gaussian (QBG) classifier. Five different types of motions are selected for this research which are: bending, gun shot, jumping, kicking and punching.