Kernel Correlation Filtering Algorithm Based On A Dual-Feature Model

Object tracking plays an important role in the field of computer,which has a wide range of applications,such as intelligent transportation,industrial detection,medical treatment and so on.Object tracking is to determine the trajectory and scale of the target in the video sequence after the first frame of the video sequence is given the location and size of the target in the absence of prior knowledge of the target.In recent years,mainstream tracking algorithms are mainly divided into two categories:Target tracking algorithm based on deep learning and using deep convolution features is highly accurate,but it cannot be tracked in real time nor applied to actual situations.The tracking algorithm based on correlation filtering uses HOG(history of oriented gradients),CN(color name)and color histogram as features,which is fast,but its accuracy is poor.To balance the real-time tracking capability and accuracy of the target tracking algorithm,this study proposes a dual-model kernel correlation filtering algorithm based on the combination of the accuracy of the deep convolution feature algorithm and the speed of the correlation filtering algorithm.The main innovations of the algorithm in this paper are as follows:(1)Dual model adaptive mechanism.The dual model mechanism includes a main feature model and an auxiliary feature model,and two of them cooperate with each other.The main feature model using shallow texture information features(HOG features)has low dimension and high computing speed.For the image sequence with clear texture features,the main feature model is used for tracking,and updated in real time.When the target position is occluded or non-rigid deformation in the video sequence,making the confidence response value of the main feature model become lower,the auxiliary feature model with deep convolution features is called.And when the confidence response value returns to the normal range,the main feature model is called again for real-time tracking.(2)Dimensionality reduction of high-dimensional features using principal component analysis techniques in an auxiliary model.The auxiliary feature model adopts the convolution feature based on the deep convolutional neural network(conv5-4 layer of the VGG-19 network and the feature dimension is D=512),so the high-dimensional feature will lead to the reduction of the computational efficiency of the algorithm.In order to further improve the speed of the algorithm and ensure the real-time performance of the algorithm,principal component analysis(PCA)is used to reduce the dimension of high-dimensional deep convolution features.On the premise of maintaining the validity of the original feature,the original feature is reduced from 512 dimensions to 40 dimensions,improving the running speed of the algorithm.(3)The main feature model is based on the KCF target tracking algorithm.The scale optimization strategy is added to the KCF algorithm to improve the performance of the tracking algorithm.The KCF algorithm finds the position of the maximum value in the response through the global search method,and further determines the target position.In this paper,the algorithm of Newton iteration is used to solve the optimal solution position of the target.We compare our algorithm with current advanced tracking algorithms with real-time speed,such as SiamFC,MEEM,SAMF,DSST.The OPE result of the public dataset OTB-2013 shows that the proposed algorithm ranks first in terms of distance precision rate and second in terms of overlapping success rate.The average speed of the proposed algorithm reaches 38fps,which satisfies the real-time requirements of the algorithm.