Abstract: Considering that the Tracking-Learning-Detection (TLD) framework can hardly balance the tracking of the whole and the interest,by using region-marking method of single bounding box,this paper proposed an improved algorithm introducing a region-marking method of double bounding boxes．With one bounding box labeling the whole object,the algorithm draws another bounding box over the stable area of image to indicate the interested part．While extracting the trace points,the weighting approach is adopted to produce more points in interested area,which improves TLD’s adaptability to local variance．Experimental results show that the improved algorithm has good performance in tracking the object when a fixed part keeps stable but the rest varies．For the object not containing the stable local area,the effects are not so obvious.

Key words: TLD,Region marking,Object tracking,Bounding box