Research On Tracker Fusion Based On Manifold Ranking

As one of the important branches of computer vision,visual tracking is at the core of motion analysis,video compression and behavior recognition,and it is widely used in video surveillance,human-computer interaction and intelligent navigation.Although there have been great developments in visual tracking over the past several decades,a variety of challenges(illumination changes,occlusion,attitude changes,background clutter,etc.)are encountered during the tracking process.The efficient tracking algorithm is still a very challenging problem.So far,many researchers put forward a variety of tracking algorithm.However,the existing visual tracking algorithm can not solve all the challenges effectively,and the comprehensive performance is not good.First of all,the purpose of the algorithm based on generative model is to learn the appearance model,the effective search algorithm and the efficient matching mechanism,but these three goals are often difficult to achieve in the tracking process.Secondly,the background is often subject to drastic changes during tracking.when the background and object are very similar,so it is difficult for the discriminative model to separate the target from the background,resulting in tracking shiftt.Therefore,in view of fusion,this topic plan to use complementary mechanisms to improve the comprehensive performance of the tracker.Because manifold ranking algorithm based on graph theory can utilize the manifold structure of large-scale high-dimensional dataset efficiently,it has been widely used in image retrieval,and has shown excellent competition and flexibility.In view of the present situation,this thesis introduces manifold ranking into the visual tracking domain,and is used in the fusion of multiple tracking algorithms,in order to improve the robustness and accuracy of the tracker.The main work of this thesis is as follows:(1)The influence of features such as Haar-like,HOG and SIFT on target tracking performance is analyzed.The multi-feature fusion strategy is adopted to describe the tracking target better;(2)This thesis present a multi-feature fusion mechanism in order to represent the target more accurately.In order to reduce the computational complexity,this thesis uses the sparse matrix to compress the feature vectors;(3)This thesis present a new model updating mechanism in order to deal with model degradation problem on the track process,is proposed.The tracking result of each sub-tracker is used as the label node.When the result of the final fusion is determined,each sub-tracker is updated with the final result.In order to make better use of the results of manifold ranking,this thesis treat the first ranking results as the label data of the next frame;(4)In order to evaluate the algorithm more fairly and objectively,we use VTB datasets as test data.This thesis mainly evaluates the algorithm from two aspects:quantitative evaluation and qualitative evaluation,and compares our algorithm with the 10 algorithm.this thesis tests the temporal robustness(TRE)and spatial robustness(SRE)of each algorithm.Finally,the experimental results show that the fusion framework proposed in this thesis can improve the robustness and accuracy of the visual tracking,especially in dealing with challenges such as illumination,rotation and occlusion;...