Classifying human body motions using Gabor features

The paper describes a method for classifying the motions of human bodies in an image sequence. First, a set of templates is prepared in advance, which includes the spatio-temporal Gabor features of key motions. Next, processing is performed to obtain the Gabor features of all unknown motion. Correlation coefficients between the feature vectors of both the key motions and the unknown motions are then calculated by using dynamic programming (DP), and finally the unknown motion is classified as one of the key motions. This study also compares the effectiveness between Gabor features and principal component analysis (PCA) for sequences of postures. Experimental results using image sequences from a volleyball game show the effectiveness of the proposed method