Research On Visual Single Object Tracking

Video surveillance has an urgent demand for computer vision technology,where visual obj ect tracking is an important branch.By tracking the obj ects of interest and recording the traj ectories,a video can be compressed into a structured video summary and the abnormal behavior of group or individual can be detected.Also it can help the public security organs find the criminal suspects.After nearly two decades of development,existing tracking approaches have succeeded in tracking the objects in common video scenes.But,it still remains challenging for these approaches to deal with complex video scenes with target occlusion,deformation,fast motion,and so on.This thesis analyzes and discusses the development of visual object tracking,as well as proposing several solutions to the problems of existing tracking methods.The main contexts and contributions of the thesis are as follows:1)A long-term object tracking system based on failure detection is proposed,amining at ad-dressing the problems of unable to detect the tracking failure and easy to introduce the background information into the tracking model under occlusion in existing tracking approaches.Based on the basic model of structured support vector machine,the tracking system is extended with a track-ing failure detection strategy,which combines the log-polar transformation and Gaussian mixture model.When tracking failure occurs,the tracking model and failure detection model stop updat-ing,and the object re-detection scheme is activated to search for the object with extended tracking windows.With the help of the presented approach,the background information will not be intro-duced into the tracking model,and the tracking system has the ability to re-track the target object after the occlusion.2)A visual tracking approach via graph regularized kernel correlation filter and multi-memory voting is proposed.It focus on constructing the target manifold structure with the graph model.The relation of the training samples and proposals is mined by the score propagation on the tar-get manifold.Also the accuracy of the tracking model is increased by introducing the manifold regularization into the basic correlation filter model.The filters are obtained by solving the semi-supervised optimization problem.In order to improving the representation ability of target fea-tures and reducing the accumulation of error information on the target templates,a template pool is constructed and updated to retain several object features with different postures.Every tem-plate is given a importance weight and this weight is updated according to the usage frequency.The template with minimum weight is first replaced by a new one.In the process of tracking,several optimal templates are automatically selected to vote for the final location of the object.Benefiting from the stored rich templates and voting strategy,the proposed method can handle the problems of complex target deformations and occlusion successfully,and the tracking accuracy and robustness are improved.3)A correlation filter based object tracking approach via affine transformation prediction is proposed.It focuses on the problem of tranditional correlation filter based trackers that they cannot predict the target affine transformations except for translation.This approach adopts a coarse-to-fine tracking scheme.The update ratio of correlation filter model is adaptively generated by a Long Short-Term Memory(LSTM)network,and the coarse translation is predicted by this model.An affine estimation network is presented to estimate the coarse target posture change and translation,which are used for target alignment.Then the related features with the target affine transforma-tion are extracted by the log-polar transformation and correlation operation.Finally,several fully connected layers are connected to the features to output target affine transformation parameters.Compared with tranditional tracking methods based on the translation and scale predictions,the proposed approach transforms the tracking problem to an affine transformation parameter predic-tion problem,so as to deal with more complex target motions.