Study Of Action Recognition In Videos

Human action recognition in videos is a very hot research field,with the rapid development of camera,mobile phone and other electronic products industry,the application of human action recognition is proposed based on the increasingly high demand.For locating human action in the video,fusing multiple features extracted from the video and improving the classification effect of actions through using action lables,we proposed a new algorithm to action recognition that based on manifold metric learning.Firstly,according the human body region,the area function of human body is used for analyzing the extent of limbs stretch.Since the area function always concludes noise,we smooth it by using robust local weighted smooth method.Segment the action at the point of area function where is the minimum.Secondly,we take out a 7 × 7 covariance matrix descriptor which fuses multiple features that are temporal feature,spatial,optical flow,vorticity and divergence features which are all extracted from action segment.Finally,we find a better metric function on manifold in training stage.Based on the label information,it improves the degree of polymerization between similar point and increases the differences between different categories.The metric method improves the effect of action classification.In the experiments,the segmentation experiment results on the Weizmann public action database shows that the segmentation method proposed in this paper has strong segmentation ability.The comparison between with manifold metric learning and without when classification on Weizmann shows that the manifold metric learning enhances the effect of action recognition by 2.8%.The action recognition average success rate on Weizmann and KTH is 95.6% and 92.3%.On both Weizmann and KTH datasets,the experimental comparisons indicate that our algorithm is better than some state-of-the-art works.As experimental results show that the algorithm in the process of pretreatment action segmentation effect is ideal,describing the action to covariance matrix has good ability to fusion multi-characteristics,and the manifold metric learning method on the accuracy of action recognition is improved obviously.