Visual Tracking Based On Correlation Filter

Visual tracking is a very important research direction in the field of computer vision.It includes many cutting-edge technologies such as image processing,pattern recognition and artificial intelligence.Visual tracking is mainly used in video surveillance,human-computer interaction,unmanned aircraft,intelligent transportation,national defense construction and other fields.Tracking tasks need to deal with complex external environment and the deformation of the target itself,which always lead to the failure of tracking task.Therefore,if we could find a kind of robust tracking algorithm to settle the problems in the video,it will have important research value and broad application prospect.In recent years,the visual tracking algorithm based on the correlation filter has made great progress.The method extends the number of samples by employing the nature of the cyclic matrix,uses the target and its surrounding background area to train the classifier online,and the process of classifier training and target detection is transformed into the Fourier domain by FFT,which greatly accelerates the operation speed.Therefore,this method has high tracking efficiency and good expansibility.This paper analyzes the shortcomings of the existing correlation filter tracking algorithm and improves the tracking method accordingly.The main contribution of this paper include two points :(1)This paper proposes a tracking algorithm based on target-enhanced adaptive updating template.Firstly,this paper proposes a method to enhance the target edge feature based on the HOG feature principle which is mainly in the edge region of the target.In this way,the performance of the target is stronger.And then we analyze the problem during template updating process that the weight can't be updated following the target appearance when the current frame is changing.By computing the target response PSR size to adjust the weight of the current frame template adaptively,the template is more effective and tracking effect improved.Finally,a typical 15-segment video experiment is selected.The average distance accuracy of the method is 21.62%,the average success rate is improved by 7.55% and the average center position error is reduced by 9.6 pixels relative to the KCF tracking algorithm.Results show that our algorithm has strong adaptability in illumination change,target occlusion,rotational deformation,fast motion and so on.(2)A multi-scale target tracking algorithm based on adaptive feature fusion is proposed.Firstly,there is a problem in the traditional tracking algorithm.The feature fusion can't be changed according to the different characteristics of the image.In this paper,we proposed an adaptive feature fusion method by introducing the color information entropy to measure the color information contained in the image,so as to adapt to the change of color feature weight.Secondly,from the point of the target scale update,some algorithms choose the optimal scale size by constructing the scale gold tower method.To deal with the problem of miscalculating the target scale size only with the detection result of one frame.In this paper,through constructing the Bayesian estimation model,we construct the Gaussian function with the previous frame scale as the prior probability and the current frame response is taken as the observation value.The target scale is solved by maximizing the posteriori.In this paper,we do experiments on the benchmark2013 and benchmark 2015 database,the accuracy and success rate were significantly improved by using our method.