Visual Tracking Algorithm Based On Correlation Filter

Object tracking is an important research branch in computer vision.It has been widely applied in many fields,such as intelligent transportation system(ITS),video surveillance,medical diagnosis and so on.Recently,the correlation filter based trackers have become one of the mainstream research methods in object tracking by virtue of its efficient computing power.And there are many parts of correlation filter,such as feature selection,sparse update,scale prediction and so on,that be innovated and improved by some scholars,resulting in an effectively improvement in tracking performance.On the other hand,for the deep learning based trackers,the tracking accuracy is excellent but the speed is slow.So they need a graphics processing unit(GPU)to speed up.Therefore,in order to deal with the complex and comprehensive tracking problems,such as deformation,occlusion,illumination variation and so on,the research improve in some terms of feature extraction,filter training,target positioning,etc.The ultimate goal is to design high-precision,high-real-time tracking algorithms on the central processing unit(CPU).The main content of the study is as follows:(1)A weighted correlation filter tracking algorithm based on context and relocation is proposed.In this algorithm,the histogram of oriented gradient and the color attribute are studied and the two-feature weighted fusion strategy is designed to improve the tracking accuracy.Then,the sample extraction and search matching of correlation filter based trakers are studied.The object is classified by its size and shape.From these,the object classification algorithm based on multi-scale search area and context feature is proposed to further improve the tracking performance.In order to reduce the failure of location,the peak to side ratio and the frame difference are used to determine a reasonable detection mechanism and carry out relocation operation,which also can improve the tracking accuracy again.Finally,on the OTB-2015 dataset,the average distance precision and the average overlap precision are 89.2% and 80.6% respectively,which are 4.2% and 2.67% higher than the ECO-HC algorithm.In addition,the tracking speed on the CPU is 65.2 frames per second.(2)A visual tracking algorithm based on anisotropic Gaussian distribution is proposed.The algorithm analyzes the training process of correlation filter and the feature distribution of square and rectangle.And then,the anisotropic Gaussian function is put forward to calculate the expected output of target,which improves the tracking accuracy.In addition,the statistical color histogram feature is introduced to predict the target position,and then the position will be weighted fusion with the other position which is obtained by the histogram of oriented gradient and the color attribute.The strategy further improves the tracking accuracy.Finally,on the OTB-2015 dataset,the average distance accuracy is 89.6% and the average overlap accuracy is 83.7%,which are about 4.67% and 6.62% higher than the ECO-HC algorithm,respectively.And the tracking speed on CPU is 42.6 frames per second.For the above two improved algorithms,the tracking performance can be improved to varying degrees,and the tracking speed can be much higher than the requirement of real-time tracking.The weighted correlation filter tracking algorithm based on context and relocation places emphasis on the application of target appearance and texture information to realize the personalized tracking requirements of different objects.The visual tracking algorithm based on anisotropic Gaussian distribution focuses on improving the characterization ability of features and the accuracy of filter,which fundamentally improves the overall tracking performance.