Research On Object Tracking In Plane Rotation Based On TLD Algorithm

Object tracking in video sequences is an important topic in the field of machine vision research and has been widely applied in various areas.However,achieving accurate object tracking is still facing a huge challenge because of the complexity of scence and the uncertainty of object movement.When plane rotation occurs which leads to pose variation and shape deformation in the process of tracking,most of the existing algorithms are difficult to track the target accurately,resulting in tracking drift.Especially,when there is plane rotation and the object in motion at the same time,current tracking algorithms are unable to track the target.The article researches the object tracking in the plane rotation and concrete contents are as follows:The P-N learning mechanism studies the labels of samples according to the results of tracking module and detection module,and then revises positive and negative samples to react the tracking module and detection module in the TLD(tracking-learning-detection)algorithm.The P-N learning mechanism has a good effect in the process of tracking.Therefore,the article studies the implementation principle of each module in the TLD algorithm.And then,make the overall structure of the TLD algorithm and its P-N learning mechanism as the foundation of the proposed algorithm.Affine transformation can rotate the position of coordinate system or bounding box of the target according to the different freedom parameters.Therefore,the proposed algorithm introduces the affine transformation into the tracking module for the plane rotation during tracking.When there is plane rotation during the object tracking,using the affine transformation to achieve accurate tracking.Sparse prototype can deal with occlusions which are composed of PCA base vector and trivial templates.In order to solve the problem of plane rotation and to make the algorithm more suitable,the sparse prototype target tracking algorighm SP-TLD(sparse prototype-TLD)with affine transformation is proposed based on the idea of TLD algorithm.During the object tracking,a frame image produces a large number of scan sub-windows and many of them do not contains the object of interest.In order to reduce the input of unnecessary sub-windows,the proposed algorithm introduces Kalman filtering algorithm into the detection module.To solve the tracking problem caused by the plane rotation,the article put forward the Kalman filtering algorithm with the adaptive factor.Through automatically adjusting the parameters of the Kalman filter to realize the optimal estimate.Before the detection,Kalman filter is used to predict the approximate area where the target may exist,and just input the sub-windows which have intersection with the region to the detection module.Thereby the proposed algorithm can improve the detection efficiency reducing the number of input sub-windows.So the article proposes the SPK-TLD(sparse prototype-Kalman TLD)object tracking algorithm based on the idea of TLD algorithm.Finally,comparing with the TLD algorithm and IVT algorithm which can deal with the plane rotation,the experiments demonstrate that the proposed tracking algorithm are more accurate and has better applicability,robustness and efficiency.