Research On Scale-adaptive Visual Object Tracking Algorithm

Object tracking technology has been the focus of computer vision and image processing.It has been widely used in military and civil fields,especially in video surveillance,human-computer interaction,robot and so on.A large number of scholars have been engaged in the research of object tracking,but due to the variability of target information observed in the tracking process,the maneuverability of the target and the complexity of the background,and so on.Object tracking is still a very challenging problem.In particular,the scale of the target changes when it moves in a video frame,if the size of the target template is not updated in time,then the tracking target will be matched with the wrong target template,which will cause the tracking error.Scale-adaptive ability is a necessary condition for a robust tracking algorithm.In this paper,target tracking algorithm with scale-adaptive ability is studied,the main work includes the following two aspects:(1)We proposed a scale-adaptive object tracking algorithm combining foreground discrimination and circular search,in which image perceptual hashing is used to describe and match target,and two tracking strategies are used in the tracking process.Firstly,the tracking strategy(circular search)is used to search the optimal target matching position of the current frame,then the foreground discriminant algorithm is used to search the global optimal matching foreground of the current frame,and the real time scale of the target is calculated.Finally,we choose those with higher similarity with the target template as the tracking results,and judge whether the target template is updated according to the matching threshold.Experiments show that the proposed algorithm is better in accuracy,accuracy and real-time than the classical MeanShift algorithm.It has an excellent result when the target is not moving fast.(2)We also proposed a scale-adaptive tracking algorithm based on correlation filtering and perceptual hashing.Firstly,the kernel ridge regression is used to minimize the mean square error between the training image and the target.A robust filter template is constructed to track the central position of the target.Secondly,by tracking the four sub-blocks of the target image,the length and width scaling coefficients are calculated to update the target scale.Finally,the target template is updated using perceptual hash technology.Experiments show that the proposed algorithm is better in qualitative and quantitative analysis than the other 10 mainstream algorithms in the OTB database.