Research On Visual Tracking Algorithms Based On Correlation Filters

Visual object tracking algorithm based on correlation filter has attracted wide attention due to its high tracking accuracy and fast computation speed.However,the traditional tracking algorithm based on correlation filter has defects in principle,and can only achieve discrete location in pixels,which reduces the tracking accuracy.At the same time,the scale estimation module of correlation tracking relies heavily on the accuracy of target location,and it can not re-detect the target after heavy occlusion.In order to solve these problems,this thesis propose a long-term tracking algorithm based on continous correlation filters which could achieve sub-pixel target location.The algorithm constructs a convolution operator for the target location in the continuous domain,and formulates target re-detection algorithms to re-detect the lost target after heavy occlusion.We evaluate the proposed algorithm and the results show that our method could achieve very high tracking accuracy.In this thesis,a correlation filter tracking algorithm in continuous domain is first studied.The algorithm divides the visual tracking task into two parts,which include target location and scale estimation.In the target location module,the algorithm exploits discriminative features from the convolution neural network and designs a convolution operator to continuous locate the target.In this way,the propose algorithm effectively addresses the target drift problem caused by discrete object positioning enhances the target location ability compared with traditional tracking method.In order to solve the problem of incorrect scale estimation when exist light location error,we modify the traditional model and present a novel multi-pyramid strategy based on HOG feature,which realizes robust scale estimation and improve the tracking performance.When tracking the object in long-term videos,if it comes across serious occlusion,the tracker will lose the target and result in tracking failure.This is a challenge task in visual tracking.Aiming at solving this problem,this paper proposes a target re-detection module based on Random-Ferns and Mean-Shift theory.The algorithm learns the target model in stable frames of the video(when there is no motion blur and target occlusion).When the target get lost,the algorithm will search for the target in a wide range centered at the disappearing position.We evaluate the these methods and the results demonstrate the effectiveness of the re-detection algorithm,which enhance the stability of the tracking algorithm.It is of great important to thorough evaluate visual tracking algorithm.And a special video library and evaluation method are necessary to evaluate the tracking algorithm comprehensively and objectively.This thesis studies the visual tracking benchmark of OTB-100 and VOT2015,and use different evaluation indexes to measure the performance in different aspect.We conduct numerous experiments in the two datasets respectively.The experimental results show that the proposed algorithm can solve the tracking tasks in different scenes very well,and has a strong competitiveness compared with another state-of-the-art tracking methods.