Human Behavior Recognition Based On Video

With the development of computer and Internet,human behavior recognition based on computer vision plays an important role in life and production.At present,the domestic and foreign research in this field mainly aims at the innovation and improvement of feature extraction and classifier design,especially in the feature extraction,including global feature(such as motion energy image,motion history map),local feature(such as Harris-3D,Hessian,HOG/HOF).Recently,the behavior recognition method based on trajectory feature had shown that it outperformed the state of the art,which are intensively sampled in the optical flow field and the sampling points are described to identify human behavior.But because of the background and motion related areas difficult to distinguish,especially in the presence of camera motion,which leads to the lack of the validity of the motion characteristics and the trajectory expression,this paper has done the following work:(1)In the process of feature extraction about saliency trajectory,considering action-related regions are changeable but continuous with time,we temporally split a video into sub-videos and compute the salient regions sub-video by sub-video.In addition,to ensure spatial continuity,we spatially divide a sub-video into patches.Constructing motion matrix of video using optical flow information,based on the observation that the motion in action-related areas is usually much more irregular than the camera motion in background in the situation of camera motion.Thus,implementing low-rank matrix decomposition on the motion information in each frame is a suitable way to discriminate action-related areas from background in the view of the relative motion of the camera,and then motion saliency region is obtained.(2)A video background region is included in the motion dependent region detected by the above method in the complex scene in the video.In order to remove the pixels in the background region,a complete moving object template is proposed by combining the edge detection and the background subtraction method,and then the saliency region is obtained.(3)In the process of tracking,median smoothing filter is used to remove the extreme pixels,And then sampled in the optical flow field and determine whether it is located in the salient region to get salient feature points.Finally,the iterative tracking is accomplished by a dense trajectory tracking method.