The common template processing method is difficult to adapt the con stant changing of target and background in visual tracking system.Therefore,an auto-reconstructing object tr acking algorithm based on multi-level dictionary is presented to overcome these shortages in this paper.The proposed algorithm can enhance the capacity to describe the target in time and space by building multi-level dictionary,which contains local details and whole information.The template can be split and merged according to foreground and ba ckground by using multi-level dictionary in the tracking progress.The split process can generate more tracker s and improve the precision of tracking effectively from different directions.On the other hand,the merging process can reduce the computational load.By qualitative and quantitative analysis,experimental results show that the propos ed method can gain good accuracy and efficiency,and is robust to deformation and occlusion.