Features extraction approach based on dense salient trajectories in videos

Motivated by dense sampling technique and researches on saliency, a novel human action recognition approach based on adaptively-fused-saliency trajectories in the video are presented. First, trajectories are extracted, and subsequently appearance and motion saliency analyses are performed to capture the static and dynamic visual attentions respectively. Both saliency values are adaptively fused as a saliency map to incorporate complementary visual information and concurrently overcome the problem due to camera motion. Then, salient trajectories are purified by taking the advantage of saliency maps. Thirdly, video sequences are transformed into the vector space of visual words through the BoVW model. Lastly, a nonlinear SVM classifier is employed for classification. Experiments on KTH and UCF datasets validate the effectiveness of the proposed method. Experimental results demonstrate that our method achieves comparable results with the state of the art.