| As a fundamental approach in computer vision,image alignment method provide better input for other video analysis approaches to boost their performance or increase the robust-ness.With the wild spread of various types of cameras,the amount of image data increases exponentially.Meanwhile,image data captured from portable cameras often suffers from blurry or shaking caused by camera motion.Direct extract high level information from these contaminated image data is troublesome.Under such circumstance,on-line robust image alignment is the proper approach to align massive contaminated images which pro-vides stable input for followup computer vision approaches.By simultaneously estimating the underlying structure of images and the corresponding geometric transformation,on-line robust image alignment can align images effectively and efficiently.Background subtraction method via online image alignment is capable of dealing with camera motions.By robustly aligning the background,the proposed background subtraction approach can achieve bet-ter performance.Visual object tracking approach based on image alignment can accurately evaluate the location of the target despite its varying appearance.Beside,target location re-estimation scheme was proposed to alleviate the drifting phenomena during the tracking procedure,which leads to a robust tracking result.The conventional image alignment methods usually take bach optimization approach to align images which require all the observation to estimate the geometric transformation between images.To solve this problem we propose a novel online robust principal component analysis method.By assuming the matrix constructed by images contaminated by sparse error have relatively low rank,online robust principal component analysis simultaneously estimates the noise and updates the subspace.In this thesis,we propose new approaches and techniques on online robust image alignment,robust background subtraction and robust visual object tracking.The main contribution can be summarized as below.First of all,image alignment plays important role as a preprocess procedure in computer vision applications.Since real world image sample often suffer from occlusion,illumination variance etc.,online robust image alignment is able to distinguish the noise from the sample thus extract better extract the template and further align the images.By proposing two novel basis update techniques,the subspace can be better extracted.Basis update techniques base on close-form solution enjoys faster extraction of the subspace while stochastic approaches recover the subspace better.Experiments on both synthetic and real world data demonstrate the effect of the proposed method.Second,background subtraction approach extracts target of interest from the back-ground.Different from traditional approaches,we propose a background subtraction ap-proach which can simultaneously distinguish foreground object from background and esti-mate the background transformation.By evaluating the motion between frames,background can be better modeled.Moreover,a combined basis update technique which enjoys both the fast convergence property and better recover result is proposed.Experiment on real world data shows the performance of proposed background subtraction approach.Third,tracking the target during the image sequence is a fundamental process for com-puter vision approaches.In real world image sequences,targets often suffer from illumination variance,occlusion,rotation,etc.,which significantly influence the target appearance.To improve the robustness of the tracking approaches,we propose an image alignment based tracker.By granting the ability of removing errors introduced by occlusion,deformation,etc.,to particles,the proposed particle filter based tracker can accurately estimate the loca-tion of the target while better modeling the appearance.Moreover,inverse composition was introduced to reduce the computation requirement.Experiment on challenging sequences demonstrates the effectiveness of proposed tracker.Finally,the commonly observed drifting phenomenon often degrade the performance of a tracker.Drifting phenomenon is caused by accumulated small errors introduced during the tracking procedure.To alleviate the drifting problem,we proposed a temporal-adjust cor-relation filter based tracking approach.By introducing a bundle adjustment like approach,which re-estimates the location of the target in time window,the proposed approach can reduce the error introduced to model during tracking.Through accurately locate the tar-get,the proposed tracker outperforms other state-of-the-art trackers in real world datasets.Moreover,as a universal component,our target location re-estimation technique can be apply to other discriminative correlation filter trackers. |
