Research On Dance Action Recognition Based On Multi-Features Fusion

Action recognition is a very challenging research topic in the field of computer vision.In Recent years,because of its high research value,it has become a very hot research direction.Recently,a large number of scientific research institutes and researchers have done in-depth research on this subject and they also have achieved some great achievement.However,the research of action recognition technology combinating dance is still in the initial stage,due to the problems of high complexity between dance actions,the progress of dance recognition research based on video is relatively slow.In this paper the above-mentioned problems have been taken into account,and the research,analysis of current areas of action recognition have been done.Aiming at the combination of action recognition and dancing video,this paper mainly focuses on feature extraction,representation and action recognition based on dance video.The detailed research content and innovations include the following three aspects:(1)First,methods of action recognition is classified.Second,this thesis mainly studies the extraction and representation of features in action video,which includes the Histogram of Oriented Gradient feature,Histogram of Optical Flow feature,BoW model and audio features involved in this thesis.(2)An effective feature extraction method for dance video is presented in this thesis.About this method,first,the dance action video of dataset should be segmented into equal sub-video.And we will execute accumulate edge feature operation which accumulates the edge features of every images in the sub-video into an image at last.Then we extract the histogram of oriented gradient feature until the option has been finished.Finally,a set of histogram of oriented gradient feature vector is used to describe the appearance information and shape of dance action.(3)A multiple feature fusion action recognition based on histogram of oriented gradient,histogram of optical flow,audio signature feature is proposed.In order to solve the problem of heterogeneous feature fusion,multiple kernel learning method is used for the fusion of these features mentioned above.Finally,the validity of the dance action recognition algorithm has been validated on the DanceDB dataset of the University of Cyprus and the FolkDance dataset which made by my laboratory.The experimental results show that the algorithm can improve the accuracy of the simple dance actions,and the method can still ensure a certain accuracy in the face of confusing background and target,which proves the validity of the motion recognition algorithm in the dance action recognition.