Research On Long-term Target Tracking Algorithm Based On Kernel Correlation Filter

Object tracking is one of the research hotspots in computer vision.It is widely used for numerous applications,such as intelligent video surveillance,human-computer interaction,etc.During tracking,object may suffer from illumination changes,scale changes,partial occlusion,full occlusion and out-of-view.Designing a tracking algorithm that can adapt to all the complex environments is a very challenging task.This paper mainly focuses on the long-term object tracking algorithms with Kemelized Correlation Filters,and makes an in-depth study of object disappearance and drift problems which are prone to appear during long-term tracking.The main work and innovation of this paper are as follows:According to the different methods of feature extraction,the object tracking is divided into traditional object tracking and deep learning based object tracking,and three classic tracking algorithms are introduced respectively.Meanwhile,all the datasets and evaluation criteria used in this paper are introduced in detail.Experiments are conducted on the six classical tracking algorithms and their performances are analyzed according to different datasets.This paper proposes an adaptive correlation filters based long-term tracker ACFLT.ACFLT utilizes an online S VM detector to re-detect the object when the correlation filter tracker failed.This helps to solve the problem of object disappearance and drift.In addition,ACFLT uses a normalized peak value to measure the confidence of the tracking results and estimate the status of the object,and then updates both the tracker and detector adaptively according to the different states of the object.Experimental results show that ACFLT outperforms other trackers on the long-term tracking dataset and also performs well on two general datasets.A long-term tracking algorithm CCFLT is proposed,which combines the deep convolutional features and shallow convolutional features to obtain both semantic and discriminant information.The combined features are used in the correlation filter based tracker for object tracking.CCFLT also uses the normalized peak value to estimate the object state and updates the model adaptively,which improves the robustness of the algorithm in terms of object disappearance and drift.On the basis of CCFLT,a long-term tracking algorithm HCFLT with hierarchical convolution features is proposed.HCFLT uses a correlation filter with deep convolution features as a base tracker,and CCFLT as an auxiliary tracker.When the tracking failure detection scheme determines that the base tracker has failed,the object location is updated considering the tracking results of CCFLT.Experimental results show that this method can improve the accuracy of tracking results effectively.