Human action recognition in videos based on dense trajectory selection

Human action recognition in wild scene is discussed and a novel approach of dense trajectory selection is addressed in this paper, in order to deal with side effects ascribed to clutter, unsteady background interference, camera motion and random noise in the video. First, multi-scale temporal pyramid is constructed from original frames in the video. By employing dense sampling, candidate initial points for dense trajectory extraction are attained. Secondly, both optical flow and median filtering are utilized to extract dense trajectories. Subsequently, both salience measurement and cosine similarity analysis are adopted in our proposed selection strategy to preserve salient trajectories. Thirdly, for video representation propose, Bag of Words (BoW) model is employed to get the corresponding feature vectors. Finally a nonlinear support vector machine is used for action classification in popular datasets KTH and ADL. Experimental results show that the proposed method achieves comparative state-of-the-art performance.