Video Motion Detection Based On Sparse Coding And Length Posterior Probability

Human action detection for untrimmed long videos is an important part of computer vision,and it plays a significant role in the security and surveillance applications.Human action detection not only the action classification,but also the action localization.In recent years,with the development of computer vision and machine learning theory,action detection methods have made many breakthroughs.However,there are still some problems in the existing methods,such as the lower proportion of real actions in action proposals,and the inaccuracy of action proposals classification.Therefore,this paper proposes an action detection method based on sparse coding and posterior probability of length,which focuses on the following aspects of research:Firstly,we propose a generate action proposals method based on temporal context sparse coding to improve the quality of action proposals.This method not only use trimmed action segment in learning dictionary,also add some information including action temporal context,to strengthen the dictionary generalization ability,thus under the same quantity generate proposals that contains more real actions.The less the number of action proposals,the lower the cost of the follow-up operation.Secondly,we propose an action detection method based on the posterior probability of length to improve the action detection accuracy.We have observed that some action proposals only a small part of the real action,but because it does not contain the background,may get a higher score in classification,so then suppressed those proposals that has lower classification score but higher overlap rate of real action in non-maximum suppression.In view of this phenomenon,after classify proposals by classifiers,we propose to adjust the classification result by using posterior probability of length,to improve the accuracy of action detection.Finally,we design and implement an action detection demo system.Users can input videos by the graphical interface,after the server processing,return the action detection results on the interface,and give the location and action category of the detected actions in videos.This system can make users feel intuitively on the result of action detection.