D338 trajectory feature is combined with kernel-based nonlinear representor (KNR) for video-based human action recognition. Initially, feature points are filtered through the Shi-Tomasi discrimination criterion after uniform sampling. Then dense trajectory is formed by tracking features based on optical flow field. For generating features of the same dimension, Fisher vectors (FVs) are used to encode features of different dimensions. Finally, these aligned FVs are taken as features for KNR training and classification. To verify the performance of combination, Gaussian kernel with the same parameter is applied to KNR and support vector machine (SVM). Experimental results on the KTH action database show that the proposed method can significantly improve classification performance with respect to SVM.
展开▼
