An improved method using kinematic features for action recognition

Human action recognition is a challenge problem in computer vision. In this paper, we propose an improved approach using kinematic features for action recognition. In this approach, we find the area that relates to action by a simple method, and select eight discriminative features derived from optical flow field to describe the dynamics of the field. The covariance matrix of the feature vectors is used to fuse the features and to serve as the feature descriptor. Multi-class SVM classifiers are then employed for action classification. Experiments are carried out on public datasets. We obtain a recognition rate of 97.66% SEG-ACA and 98.2% SEQ-ACA on KTH dataset, and 98.89% SEQ-ACA and 93.83% SEG-ACA on WEIZMANN dataset with leave-one-out test.