Visual Tracking Method Based On Sparse Representation

Visual tracking estimates the object position at each time in the obtained video or image sequence.It is widely applied in the fields of behavior analysis,video surveillance,autopilot and human-computer interaction.In the past decades,some remarkable progress has been achieved in the field of visual tracking,and numerous tracking algorithms with high efficiency and robustness have been proposed.However,some challenging problems,for example,illumination variation,scale variation,occlusion and background clutter,have not been solved effectively,which leads to the significant degradation of tracking performance.Therefore,Visual tracking is still a challenging task.In order to improve the accuracy and robustness of visual object tracking algorithm,this paper studies this problem based on sparse representation theory.The main research work and innovations are as follows:Focusing on the issue of heavy decrease of object tracking performance induced by illumination variation,a visual tracking method via jointly optimizing the illumination compensation and sparse representation is proposed.The template illumination is firstly compensated by the developed algorithm,which is based on the average brightness difference between templates and candidates.In what follows,the candidate set is exploited to sparsely represent the templates after illumination compensation.Subsequently,the obtained multiple optimization issues associated with single template can be recast as a multi-task optimization one related to multiple templates,afterwards the similarity between the candidates is considered to implement low rank constraint on the sparse coding matrix,and the sparse error term is included to improve the robustness of the algorithm to local occlusion,thereby an illumination compensation and sparse representation joint optimization model can be constructed,which can be solved by the alternative iteration approach to acquire the optimal illumination compensation coefficient and the sparse coding matrix.Finally,the obtained sparse coding matrix can be exploited to quickly eliminate the unrelated candidates,and then the local structured evaluation method is employed to achieve the object tracking with high accuracy.Focusing on the issue of heavy decrease of object tracking performance induced by complex background and occlusion,a visual tracking method via discriminative dictionary learning is proposed.The object and background samples are firstly obtained according to the local correlation of the object in temporal-spatial domain.In what follows,a dictionary learning model is established based on sparse representation: the outliers generated by occlusion are captured by error terms,and the sparse encoding matrix and error matrix are punished by nonconvex minimax concave plus(MCP)functions,in addition,inconsistent constraints are imposed on the dictionaries to improve the robustness and discriminability of dictionaries.Concerning the established nonconvex dictionary learning optimization issue,the majorization-minimization inexact augmented Lagrange multiplier(MM-IALM)optimization method can be exploited to get better convergence.Finally,the reconstruction errors of candidate object are computed from the learned discriminative dictionary to construct the object observation model,after that,the object tracking is realized accurately based on the Bayesian inference framework.In summary,this paper deeply studies the visual object tracking algorithm based on sparse representation,and carries out experimental analysis.The simulation results show that the proposed method can improve the accuracy and robustness of the object tracking significantly under the challenges of illumination variation,scale variation,occlusion and background clutter.