Research On Kernelized Correlation Filter Tracking Algorithm And Its Improvement

The key technology of precision-guided weapons is the accurate guidance technology,and image object tracking is a typical application of accurate guidance technology.How to achieve fast and stable image object tracking has always been important and difficult problem in the field of image-guided technology research.This paper did some research on real-time object tracking in complex background.The main work includes the following sections:First part of this paper is research on kernelized correlation filter,including model training,object detection and dense sampling method based on circulant matrix.To tackle the shortness that KCF cannot predict target scale,this paper proposed scale pyramid searching strategy.The tracker will be extended to two stages,the first stage output the two-dimensional coordinate position of target,and in the second stage,this algorithm will build the scale pyramid to determine the best target scale.Experiments showed that this algorithm can effectively solve the target scale variation problem.The second part made two improvements of KCF to face the challenges for a long term object tracking task.Firstly,this paper introduced the initial object model when updating target model.And this strategy can effectively suppress model drift.Secondly,this paper added an object detect module.When the tracking module fails,the algorithm will be switched to the object detection module.Target detection module uses two detectors,the first detector is variance classifier,and the second one is KCF.The last part of this paper is consist of object tracking experiments and result analysis.Firstly this paper introduced a tracking algorithm evaluation system,including evaluation methodology and test video set.Then,this paper verified that the proposed algorithm outperforms KCF tracking algorithm.Finally,this paper used evaluation system to make an objective comparison evaluation between the proposed algorithm and other outstanding tracking algorithms,and it turned out that the proposed algorithm is better than others no matter in the view of precision or robustness.