Research On Online Object Tracking Based On Sparse Representation

The goal of object tracking mainly lies in target state estimation in videos or image sequences,where the state includes location, scale, direction, et al. As a fundamental problem, object tracking plays a critical role and enjoys many realistic applications. Though much progress has been made during the past decades, developing a robust tracking algorithm is still a challenging problem. This is due to the visual difficulties caused by target appearance variations in the scenes. Factors may include varying illumination, camera motion, occlusion, scale changes, background cluttering, pose variation, image blurring, drastic movement, et al. Recently, online object tracking has gradually been popular in research area. It includes a set of new tracking methods, which take attempts to adapt and update the target appearance model during the tracking process, and is likely to obtain more satisfactory performances compared with previous methods. Especially, online object tracking based on sparse representation has been widespread concerned based on the recent theoretical success in signal representation, convex optimization, compressive sensing, et al.In this thesis, we focus on robust online object tracking based on sparse representation under different target appearance variation conditions. Main work and contributions are as follows.(1) An online sparse-representation-based object tracking method based on generative model is proposed. The target is represented with overlapped and selected local patches based on Key Point Proportion Ranking(KPPR), and its location is estimated by spatio-temporal analysis. Temporally, a propagated affine warping dynamical model is newly introduced. Spatially, observation model based on weighted sparse representation based on Elastic Net regularization and spatial confidence inference in Coordinate-Confidence Space(CCS) is newly established. Both selection pattern and templates are periodically and sparsely updated to adapt the target’s appearance variation.The proposed method can stably and robustly follow the target in 14 out of 16 evaluation sequences and ranks first on both average center error(ACE) and average overlap rate(AOR) in 1 sequence among 24 evaluation methods, where the ACE average of 12 sequences reduces below 10 pixels,and the AOR average reaches 66.8%. Both qualitative and quantitative experiments demonstrate that the proposed approach achieves more favorable performance compared with classical works.(2) An online dictionary-learning-combined object tracking method based on discriminative model, called IDSDL-VT, is proposed, which aims to solve the performance improvement issue of the proposed generative method caused by processing design without consideration on the target surroundings influence. Firstly, an dictionary learning algorithm called Incremental Discriminative Structured Dictionary Learning is presented. In our framework, a discriminative dictionary combining both positive, negative and trivial patches is designed to sparsely represent the overlapped target patches. Then, a local update(LU) strategy is proposed for sparse coefficient learning. To formulate the training and classification process, a multiple linear classifier group based on KCombined Voting(KCV) function is proposed. As the dictionary evolves, the models are also trained to timely adapt the target appearance variation. The proposed method can stably and robustly follow the target in 14 out of 16 evaluation sequences and ranks first on both ACE and AOR in 3 sequence among 24 methods, where the ACE average of 13 sequences reduces below 10 pixels, and the AOR average reaches 67.3%. Qualitative and quantitative evaluations compared with state-of-the-art methods demonstrate that the proposed tracking algorithm can solve the issue on design without surroundings consideration, and achieves more favorable performance.(3) An online sparse-representation-based object tracking method based on Hybrid Incremental Sparse Model(HISM) is proposed, which aims to expand the application scope of the proposed discriminative method and improve the efficiency of its dictionary learning algorithm.The target is represented with overlapped patches, and the tracking process is conducted in a cascaded generative-discriminative manner. After the candidates are sampled by affine warping, the algorithm generatively selects the partial patches based on newly introduced Incremental Sparsitybased Spatial-temporal Contribution Ranking(ISSt CR), and estimates the location based on the candidates voted by the patch-based multiple-linear-classifier group. The models are updated via newly proposed Residue-weighted Incremental Discriminative Structured Dictionary Learning(RIDSDL) algorithm, which is a combination of newly proposed Residue-weighted Online Dictionary Learning(RODL) algorithm and IDSDL presented above. When the index number of selected patches is variable, an adjustment algorithm based on Inter-frame Indices Difference(NAIID) is proposed. Experiments verify the efficiency improvement of proposed dictionary learning algorithms and its application in face recognition. For object tracking, the proposed method can relatively stably and robustly follow the target in all 16 evaluation sequences and ranks first on both ACE and AOR in 1 sequence among 24 methods, where the ACE average of 13 sequences reduces below 10 pixels, and the AOR average reaches 71.6%. Experiments demonstrate that the proposed approach archives a more universal application scope, comprehensively more favorable performance despite limited negative influence on tracking accuracy compared with the methods proposed above, and competitive performance compared with classical works.