Human Action Prediction In The Video

At present,the action prediction methods based on optical flow has attracted extensive attention from researchers.Usually,optical flow is suitable for action prediction with large motion displacement,but optical flow is susceptible to external factors and easy to introduce redundant information.The pose-based action prediction method is more suitable for the prediction of subtle action,but the human body pose is susceptible to the influence of factors such as illumination and occlusion,which has not received extensive attention from researchers.Therefore,this paper studies the action prediction methods based on optical flow and human pose.For action with large motion displacement,this paper proposes a deep learning action prediction method based on optical flow.In order to evaluate the redundant information in optical flow,a method is proposed for measuring the redundant information of optical flow.For the redundant information introduced by the background,the redundant information can be eliminated by selecting the target area.For the redundant information introduced by camera shake,the influence of camera shake on action prediction can be eliminated by calculating the optical flow in the background.For the redundant information in the still part of the target,the negative effect of the still part of the target on action prediction can be eliminated by detecting whether the target region is still.Based on the optical flow which are removed redundant information,an action prediction network combined spatial and temporal scales is designed.Experimental results on the UT-Interaction setl and UT-Interaction set2 datasets demonstrate the effectiveness of the proposed method.For action with small motion displacement,this paper proposes a method based on human pose and applies it to predict cheating behavior.In this method,deep learning method is used for pose estimation at first.When head pose performs abnormally,the corresponding image region is extracted,and SVM is used to determine whether it is turn back cheating.For cheating with notes or mobile phones,the hand and head pose are combined for analysis.If there is any abnormality,the images of the hand area were extracted,and the convolution network was used to determine whether there is a note or mobile phone nearby the hand.This paper uses examination room monitoring video to test the pose-based action prediction method,and it can predict most of the cheating behavior.