Single Object Tracking Via Correlation Filter And Attention Mechanism

Single object visual tracking is the most basic problem in the field of visual object tracking.The main task of single object tracking is to locate a given target from video sequence and estimate the scale of the target in order to accomplish more advanced tasks.The difficulty of tracking task is that the training data is scare and unbalanced.The tracking model learnt from the limited imbalanced training data is difficult to cope with the challenges of occlusion,deformation and scale variation in actual scenarios.Meanwhile,the requirement of the high efficiency of the visual tracking task further enhances the difficulty of the algorithm design.Tracking algorithm based on correlation filtering uses circulant shift to generate training samples.It applies the fact that cyclic matrix can be transformed to diagonal matrix in frequency domain and FFT to train the tracker and detect the object efficiently.It has gained an increasing attention in the tracking community due to their impressing performance.However,the boundary effect and the limited discriminate ability of the tracking model,further improvement of correlation filter based tracking algorithms is limited.Employing convolutional regression network to approximate the solution to the ridge regression problem will help to solve these problems.Two convolution regression algorithms based on attention mechanism are proposed to solve the problems of unbalanced training samples and the interference of similar objects in the background when using pre-trained classification network to extract features.The proposed algorithm is also evaluated comprehensively on the popular visual tracking benchmark.The experimental results show that the proposed algorithm achieves good tracking performance.Firstly,this paper proposed a spatial regularized kernelized correlation filter based tracking algorithm.In this paper,we introduce the spatial regularization component into the ridge regression model used by classical kernelized correlation filter(KCF)to improve its performance.We solve the new ridge regression formula efficiently with the property of circulant matrices.In this way,we can simultaneously keep the real-time and improve the tracking performance.On the otb-2015 benchmark,compared with KCF,the tracking accuracy rate is improved by 0.114,and the tracking success rate is improved by 0.127.Secondly,a maximum margin object tracking algorithm training with weighted circulant feature maps.Support vector machine(SVM)model is used to replace the regression model in correlation filtering.We give each training sample a weight based on their accuracy to reduce the influence of inaccurate samples.Moreover,a model update strategy is introduced to prevent the tracking models from being polluted by wrong samples.On the otb-2015 benchmark,compared with KCF,the tracking accuracy rate is improved by 0.125,and the tracking success rate is improved by 0.128.Thirdly,a Residual Attention Convolutional Network for Online Visual Tracking is proposed.Convolutional regression tracking reformulates the DCFs as a one layer convolutional network and avoids the boundary effects.We introduce a residual attention module to the one layer convolutional network to inhibit the descent of discriminative ability caused by data imbalance.And two types of activation function are applied to capture the spatial attention and time attention.On the otb-2015 benchmark,compared with SRDCF,the tracking accuracy rate is improved by 0.056,and the tracking success rate is improved by 0.028.Finally,an online visual tracking algorithm via dual residual attention learning is proposed.Using a pre-trained generic deep learned features is prone to suffer from the interference of similar objects in the background.The feature awareness residual attention learning(FARAL)learns a soft mask online to select and enhance the representative features for objects in different videos.The background awareness residual attention learning(BARAL)generates the dense samples and response maps.The residual branch learns the structure information of the target and the background around the target to guide the response maps.On the otb-2015 benchmark,compared with CREST,the tracking accuracy rate is improved by 0.038,and the tracking success rate is improved by 0.021.