Long-term Visual Tracking Based On Kernelized Correlation Filters

In recent years,Correlation Filter-based visual tracking algorithms have aroused increasing interests in the field of visual object tracking,and have achieved extremely compelling results in different competitions and benchmarks.But most of Correlation Filter-based visual tracking algorithms are still faced with a series of challenges such as rotation,occlusion,out of view and other factors which often leads to target tracking shift or even failure.In order to solve the problem of similar target interfere,out of view,scale-variant,rotation and occlusion in Kernelized Correlation Filters(KCF)tracking algorithm,this paper proposes a long-term tracking approach based on KCF,the main results are as follows:(1)To solve the poor tracking performance problem in KCF tracking algorithm when the target undergoes similar target interfere,out of view and scale-variant,this paper proposes a long-term visual tracking based on multiple features fusion of KCF.Firstly,a spatial regularization component is introduced in the learning to penalize classifier coefficients depending on their spatial location.This allows the classifier to be learned on a significantly larger set of negative training samples and uncorrupted positive samples,increasing discriminative power of learned model.Then,the Newton method was used to complete the iteration and obtain the obtain the maximizing response location and target scale score of the classifier in the detection area.Finally,to re-detect the target in the case of tracking failure and achieve long-term tracking,this paper compares the confidence of the location with maximum score and trains an online Support Vector Machine(SVM)classifier.(2)To solve the tracking drift problem in KCF tracking algorithm when the target undergoes rotation,out of view and occlusion,this paper proposes a long-term visual tracking based on adaptive convolutional features of KCF.Firstly,convolutional features are extracted from pre-trained VGG-Net convolutional neural network model.Compared to other convolutional features of VGG-Net convolutional neural network model,conv3-4 layer convolutional features have a certain high-level semantic information and rich fine-grained spatial details information.Then,based on the principal components analysis of conv3-4 layer features in VGG-Net model,the dimension of conv3-4 layer features is reduced from 256 to 130 by using adaptive dimension reduction technique.Finally,tracking model is updated by the reliable tracking results which are determined by calculating the confidence of the target position using Peak-to-Sidelobe Ratio.(3)To verify the feasibility of the proposed algorithm,100 groups of OTB-100 benchmark video sequences are tested and obtained results are compared with 43 kinds of traditional tracking algorithms and 9 kinds of tracking algorithms by CNNs.Experimental results indicate that the precision and success rate of long-term visual tracking based on multiple features fusion of KCF are respectively 0.831 and 0.624.Compared with KCF tracking algorithm,long-term visual tracking based on multiple features fusion of KCF approach respectively improves 19.9% and 31% in the precision and the success rate.Besides,the precision and success rate of long-term visual tracking based on adaptive convolutional features of KCF are respectively 0.805 and 0.608.Compared with KCF tracking algorithm,long-term visual tracking based on adaptive convolutional features of KCF approach respectively improves 16.1% and 27.7% in the precision and the success rate.Under the complex conditions such as interfere with similar target,out of view,significant scale changing,rotation and occlusion,the improved algorithms are still able to track target stably and accurately,which has very important theoretical value and applied research value.