Research On Visual Tracking Algorithms Based On Feature Fusion And Joint Decision

Visual tracking has gained extensive attention for its applications in computer vision related fields,such as intelligent video surveillance,motion analysis,human-machine interfaces,and robotics.The construction process of the visual appearance model is essential for the visual tracking problem.However,because of the complex changes in the moving target including rotation,distortion,scale variation,the background clutter,illumination change,occlusion,information loss from 3D space to 2D image,as well as real-time processing requirement,it remains a great challenge to build an efficient model with high performance.In this paper,we first focus on the construction of visual appearance model in the singletarget visual tracking problem,then deeply study the influences of feature fusion and joint decision upon visual tracking,and finally propose some new efficient algorithms to address related problems.This work is significant in theory and has a major influence on practical tasks.The main contributions are summarized as follows:1.A compressive time-space Kalman fusion(TSKF)model for visual tracking is presented.The compressive tracking(CT)method is a simple but efficient algorithm.With this method,the high-dimensional features can be mapped into a more compressive one while the most salient information is preserved.The newly proposed TSKF algorithm extends CT to the case of visible and infrared fusion tracking problem.Moreover,previous fusion trackers deal with features of different sensors individually without time-space adaptability and cannot adequately explore important information in the updating process.Unlike them,our tracking is achieved in both space and time domains in which the extended Kalman filter is applied to update and optimize the fusion coefficients.The proposed tracking approach is demonstrated to be accurate and robust in experimental results.2.A visible and infrared fusion tracking algorithm based on multi-view multi-kernel fusion(MVMKF)model is developed.In the problem of visual tracking,fusion between visible and infrared sensors can provide complementary features and consistent discrimination information between target and background.In recent years,multi-view learning received increasing attention thanks to its capability to combine different view features characterized by consistency and complementarity.The proposed algorithm considers both visible and infrared view features and applies the multi-kernel framework to learn the contribution of each view so as to build an general appearance representation according to the respective performance.Furthermore,the tracking process is accomplished with Bayes classifier in the sophisticated compressive feature level due to the significant performances of multi-view learning in both classifier and sophisticated feature levels.The experimental results demonstrate the performance of MVMKF algorithm in the respect of robustness,accuracy,and speed.3.A kernel joint tracking and recognition(KJTR)algorithm based on structured sparse representation(SSR)is proposed.Visual tracking and recognition are closely relevant and can mutually help each other.In order to achieve simultaneous tracking and recognition,the proposed method applies the optimal Bayes joint decision and estimation(JDE)model and uses a iteration process to jointly deal with decision(recognition)and estimation(tracking)which has a potential to reach the global optimization.Furthermore,our appearance model is constructed based on SSR which shows significant efficiency and effective in digging both holistic and local feature information.The contribution rate of each test sample is considered via a kernel function.Moreover,a newly proposed joint weights provide flexibility and robustness with severe appearance changes in the dynamic scene.The experimental results show that the proposed approach is able to achieve robust tracking and accurate recognition simultaneously.