Action Recognition Based on Dense Trajectories and Human Detection

Dense Trajectories have shown outstanding performance for action recognition and attract more attention in recent years. Dense trajectories are generated by tracking the points which are densely sampled, so part of the trajectories is generated by the background movement. Those trajectories not only take up computing resources but also have a bad effect on the classification. To solve this problem, a novel algorithm is proposed to remove those trajectories based on human detection. Faster R-CNN is utilized to detect the human in the frames of videos. Furthermore, Fisher vector is used to encode the features and linear kernel SVM with one-against-rest approach to classify actions. The proposed algorithm is evaluated on KTH and YouTube datasets. Experimental results show that the proposed algorithm achieves comparative performance compared with the state-of-the-art methods.