Research On Image Feature Extraction And Fusion Based On Visual Tracking

Image description is a vital part of visual tracking and different image features can affect the algorithm performance directly.Among the features,for the advantages of intuitive feature form,easy to understand and low requirement for experimental platform,traditional manual feature has achieved very excellent results in the visual tracking based on feature matching.However,how to extract image features with strong resolution and robustness and how to make multi-features adaptive fusion during the process of visual tracking are still problems to be solved.In this paper,we pursue research in two aspects of image feature extraction and multi-features adaptive fusion:(1)We proposed visual tracking algorithm based on hierarchical histogram of color name.Firstly,aiming at the disadvantage that the image features extracted by dividing the image into different layers according to the pixel value are sensitive to the change of illumination,we proposed a histogram of layered structure combining CN features.Then,designing image matching experiment to verify the effectiveness of this feature.Finally,we applied this feature to visual tracking algorithm and verified algorithm performance.The experimental results show that the visual tracking algorithm based on hierarchical histogram of color name can effectively improve the success rate of target location.(2)We proposed a vision tracking algorithm based on RPCA.First of all,analyzing the characteristics between original CN feature and dealt with PCA,and compared with the CN features based on RPCA,it is proved that the CN features processed by RPCA method have stronger resolution.After that,in two different visual tracking algorithm frameworks,representing the target with the CN features based on RPCA and reducing dimension to different layers to verify the impact on visual tracking algorithm.The experimental results show that the CN features processed by this method have stronger resolution,can better play the advantages of features and improve algorithm performance.(3)We proposed a multi-features adaptive fusion visual tracking algorithm based on color name histogram.Firstly,aiming at the problem that traditional color histogram is sensitive to illumination,combining with the CN feature which is robust to illumination,we proposed the color name histogram.Then,during the process of visual tracking,we adopted a multi-features adaptive fusion method to fuse this feature with HOG feature of the target to verify the impact on visual tracking algorithm.The experimental results prove that this method not only retains the advantages of traditional color histogram for the robustness to target deformation,but also enhances the resolution of target features in the light change.The performance of the algorithm is improved significantly.(4)We proposed a multi-features adaptive fusion method based on game theory.Firstly,in the framework of the classical vision tracking algorithm ECO,we extracted CN and HOG features of target.Then making the features play chess continuously during thetracking process,give full play to their advantages,seek their Nash equilibrium point and achieve the best balance state to realize the adaptive fusion of multiple features.The experimental results show that,compared with the traditional linear weighting algorithm using fixed coefficient for multi-features,this method can adapt well to the movement of the target in the complex background,and can effectively improve the performance of the algorithm.