Research On Key Technologies Of Visual Object Tracking

Object tracking is one of the main research area in computer vision and has received extensive attention and development.Many effective visual object tracking methods have been proposed.However,in practical applications,the performance of the tracking remains to be improved,effective and robust object tracking is still a challenging problem.In this thesis,we study online single object tracking and try to tackle challenges that present in practical tracking scenarios,including illumination variation,deformation,fast motion,occlusion and etc..We try to do some research on these issues,the main three contents are as follows:Firstly,using a single feature to train the tracking model is a drawback,which usually leads to poor robustness in object tracking.Aiming at addressing this problem,we propose a collaborative tracking model fusing hybrid features in this paper.A generative model,which uses colorbased features to compute the matching scores between target object and candidates,is combined with a discriminative model,which trains SVM with multiple features to classify the candidates as target or nontarget objects.The combination of the two models can improves the robustness and reduces drift.The experimental results demonstrate that our approach is effective for partial occlusion and background clutter in tracking,thus realizes desirable tracking performance under complex conditions.Secondly,according to the physiological characteristics of the human eye,saliency is a good way for depicting region properties but with less relative to specific target.In this thesis,we proposed a tracking method based on supervised salience detection to introduce the prior information of the target.We construct a graph based on superpixels to represent spatial information.Manifold ranking is combined with a discriminative model,which trains a classifier based on mid level cue,to extract prior knowledge of the target for the supervised random walk.The relevance between the saliency and the target can be improved through this way.We utilize integral graph method for a better efficiency during tracking.The experimental results on visual tracking benchmark demonstrate that our approach is effective for fast motion,partial occlusion and background clutter in tracking,thus realizes desirable tracking performance under complex conditions.Finally,most existing correlation filter based tracking methods focus on exploiting different characteristics with correlation filters,e.g.circulant structure,kernel trick,context information and effective feature representation.However,how to handle the scale variation and the model drift is still an open problem.In this paper,we propose a paradigm for detecting moving objects in videos by introducing motion based object proposals,which can capture the moving object effectively,and integrate the method into a correlation filter tracker to deal with the above problems.Integrating object proposals into tracking can deal with scale variation,fast motion and other problems naturally.Extensive experiments show the superiority of the proposed method.