New Methods For Dense Trajectories Selection Based On Visual-saliency Prior Model

Human behavior recognition is an important way to acquire,process and analyze information from massive video data at present.Due to the false extraction of background redundant trajectories led by complex environment,accurately extracting trajectories of foreground motion from dense trajectories is able to obtain the discriminative feature.In this dissertation,the error extraction of background-interferential trajectories under complicated conditions such as illumination change and camera movement in action recognition based on dense trajectories is focused on.The main work can be summarized as follows:A salient trajectories selecting method based on “candidate-constraint” double-layer foreground prior information to get temporal and spatial saliency of trajectories is proposed to solve the problem of large computation complexity in trajectory extraction.In the candidate layer,pixel point saliency is defined by using the temporal and spatial prior information of single point pixels to remove pseudo motion points.In the constraint layer,we use trajectory prior constraint to define trajectory saliency to remove pseudo motion trajectories.The discriminant strategy of constraint layer enables the candidate layer to reduce the discriminant condition of saliency to reduce the complexity of computing.The Bo W model based on directly sampling and clustering has the problem of unbalanced contribution among different classes to words.It's hard to distinguish the similar trajectories belonging to different action category.To make sure that each category has the same contribution to the dictionary,a three class cluster structure of “intra-class sampling,intra-class clustering and inter-class clustering” to get the dictionary is used.Experiments show that the classification rate of our method is at least 0.9% higher than other methods.A salient trajectories selecting method based on background double stream low rank decomposition with gray value and optical flow.The trajectories selection method based on foreground modeling may lose the necessary trajectories in the moving area and be sensitive to the motion of cameras.Considering the background motion caused by camera in the same time has similar characteristic,gray value and optical components from time-space video cubes to model the background motion are extracted.The saliency maps are obtained from sparse residual matrix decomposed by low rank matrix decomposition.Finally,we extract the salient trajectories of gray flow and optical flow respectively according to saliency map are extracted to get the salient trajectories set of multi view fusion.Experiment results proved that the classification rate of this method is 0.17% higher than the LFD method and 0.9% higher than the frame-difference methods.