Research On Object Tracking Method Based On Kernelized Correlation Filter

Object tracking plays an important role in the modern computer v ision.Kernelized correlation filters for visual object tracking have achieve appealing performance,but there is still a need for improving their tracking capabilities.This paper analysis this method comprehensively,and propose two schemes to improve it.To solve the tracking lost problem caused by the occlusion,a detection component is built,some of the current detection algorithms is analyzed,and the detection scheme is determined.The object template is matched with the current frame through SURF feature point,the improved RANSAC algorithm is used to filter the matching pairs to calculate the transformation matrix for locating the target position from the current frame,so the purpose of detection is achieved.A set of templates is created to increase the r obustness.A judgment mechanism is established by using the current frame to train a tracker and reverse tracking the video,then the results are compared to determine whether to start the detection,also the template updating strategy is revised.Aiming at the problem that the kernelized correlation filter can not adapt to the target scale changing,the object proposals method is introduced to generate the bounding box with different scales.The edge box algorithm for extracting the target candidate box is derived by the structured random forest,meanwhile the algorithm is modified to suit the need.The original algorithm is used for coarse tracking,the region at the result position is extracted,the edge box is performed at the region to produce the target candidate boxes with different scales,the boxes with high score are selected,the original tracking result is combined with to filter the boxes.And then they are converted to the initial size for the evaluation with KCF,the most appropriate tracking box in the current frame is obtained by integrating the original result.Meanwhile a variety of features are integrated together to further boost the overall tracking performance.This paper also discusses the tracking evaluation measures.When evaluating the two improvements,29 videos with occlusion and 28 videos with scale changing are selected from the OTB data set.Through qualitative and quantitative experiments.The proposed algorithm is compared with the original KCF,TLD,Struck algorithms.The experimental result s show that the proposed method rank over the original KCF,TLD,Struck algorithms both in the success plot and the precision plot.Compared with the original method,the proposed tracker is better applied in the case of scale variation and occlusion,and can b e widely used in target tracking...