Research On Visual Object Tracking Based On Correlation Filter

With the wide application of high performance computer and the ease of use of high quality and low cost camera equipment,which accelerate the demand for au-tomatic analysis and processing of video content.These have greatly promoted the development of visual object tracking technology in the field of computer vision.Vi-sual object tracking is of great significance and practical value in the field of computer vision.It has been widely used in real life and plays an important role in the real world.Visual object tracking only uses the initial state of the target in the first frame of the video,including the region,size and position of the target,to predict the state of the target in the subsequent video sequence.Therefore,how to track the target quickly and robustly is an urgent problem in complex scenes such as illumination change,scale change,occlusion,deformation and background clutter.This paper aims to provide people with the key technology of automatic video content analysis and processing.Based on the characteristics of video content,the cor-relation technology of visual object tracking based on correlation filter is studied.In order to improve the performance of object tracking and meet the real-time require-ments,this paper proposes a robust real-time visual object tracking method based on complementary learners.In addition,in order to improve the performance of objec-t tracking,this paper proposed a visual object tracking method based on background suppression deep features.The primary contributions of this paper are as follows:1.In this paper,a robust and real-time visual object tracking method based on com-plementary learners is proposed,which can improves the speed and accuracy of object tracking by using various low dimensional complementary features and adaptive online re-detection components.Due to the limitation of single feature representation and redundant information and noise interference of various fea-tures in complex scenarios,existing visual object tracking methods based on cor-relation filter cannot learn robust correlation filters.By explore the advantages of various features,this paper uses a variety of low dimensional complementary features to learn correlation filters,and combine with the color histogram model to improve the robustness and speed of visual object tracking.In addition,an adaptive online redetection component which is introduced in this paper,which can effectively deal with the problem of tracking failure.This paper makes a comprehensive experimental comparison on OTB2015 datasets that is widely used with the existing state-of-the-art tracking methods.The proposed method in this paper achieves the competitive performance and speed.2.A visual object tracking method based on the deep features with background sup-pression that is proposed.The correlation filters model guided by the foreground probability to focus only on the useful deep features to improve the robustness of object tracking.The existing visual object tracking methods based on deep learning usually use a certain layer in CNNs to learn correlation filters.The high-level features provide low-resolution semantic information,which cannot accurately locate the target.Since the low-level features contains rich and high-resolution texture,edge and color information,so they has poor anti-jamming ability.Considering the advantages and limitations of the high-level features and the low-level features,this paper combines the multi-layer deep features to learn correlation filter for improving the discriminative ability of the model.In order to alleviate the boundary effect,this paper uses the foreground probabili-ty map to guide the correlation filters for automatically selecting the part of the deep features that is useful to the discriminative model.In addition,a semi-adaptive model updating strategy which is proposed to avoid model pollution and improve tracking speed.This paper makes a comprehensive evaluation of the proposed method on the OTB2013,OTB2015 and the VOT 2016 challenge datasets that are widely used in the world.Compared with the existing state-of-the-art methods based on deep learning,the proposed method achieves excellent performance.Based on the research results,this paper implements a prototype system of the robust real-time visual target tracking based on complementary learners.In this sys-tem,the proposed method can track the target quickly and robustly,which verifies the technical feasibility and practicability of this method.