Influence Of Features On Performance Of Object Tracking Algorithm

Object tracking is one of the fundamental problems in computer vision with numerous applications.In recent years,various tracking algorithms have been proposed,especially based on correlation filters and depth learning methods,which greatly improve the performance of the tracking algorithm.However,there are still many problems with the object tracking technique,because the appearance of the object changes dramatically caused by motion blur,background clutter,deformation,and occlusion.How to maintain the accuracy of the object appearance model in these complex tracking scenarios is an urgent problem to be solved.The research of obj ect tracking mainly focuses on two aspects:appearance model and discriminative model.In this work,we analyze the effects of features on the tracking algorithm from the point of view of the accuracy of the target appearance model.The main work of this thesis is as follows:(1)The effects of gray features,color features and HOG features on the tracking algorithms of three classification frameworks such as logistic regression,support vector machine and kemelized correlation filters are analyzed.The experiment verifies that the features of the object is the most important to the performance of the tracking algorithm.(2)The effects of the background features of the object on the performance of the kemelized correlation filters tracking algorithm is analyzed.The experimental results show that the performance of the algorithm has a positive effect by increasing the number of background features.(3)The tracking performance of the kernelized correlation filters algorithm is improved by using convolution features extracted from VGG19 network.The influence of convolutional features of different layers on the performance of tracking algorithm is analyzed.Based on the kernelized correlation filters,the VGG19 network is used to extract the features of the object,and the effects of different convolutional layers on tracking performance are analyzed.A method of using the peak-to-sidelobe ratio to weight the response of different layers is proposed,which effectively combines the fourth layer and fifth level convolution features and achieves better results.(4)An adaptive low-dimensional variant of color name feature for Siamese network tracking framework is proposed.The principal component analysis(PCA)method is used to extract the two most significant channels of the color name features,and construct the apposite format for the siamese network with the gray features,which improves the tracking performance of the algorithm.(5)In this thesis,the benchmark set of OTB is chosen as the verification platform of algorithm performance.The effects of object features on performance of tracking algorithm is analyzed qualitatively and quantitatively.By using the convolution feature,the KCF algorithm is improved by 21.6%and 17.3%on the success rate and the accuracy rate respectively.By using the color name features,the siamese network tracking algorithm is improved the success rate and the accuracy by 4.2%and 3%.