Power difference template for action recognition

This paper proposes power difference template as a new spatial-temporal representation for action recognition. Specifically, spatial power features are first extracted according to the transform of Gaussian convolution on gradients between logarithmic and exponential domain. Using the forward–backward frame power difference method, we thus present normalized projection histogram (NPH) to characterize segmented action spatial features by normalizing histogram of the 2D horizontal–vertical projections. Furthermore, from the perspective of energy conservation, motion kinetic velocity (MKV) is introduced as a supplement for representing temporal relationships of power features by supposing that the variation of power is produced by motion in the form of kinetic energy. Our power difference template fusing NPH and MKV is further integrated to a bag of word model for training and testing under a support vector machine framework. Experiments on KTH, UCF Sports, UCF101 and HMDB datasets demonstrate the effectiveness of the proposed algorithm.