目标跟踪作为计算机视觉领域中的一个关键课题，近年来有不少优秀的算法出现，其中相关滤波类和深度学习类的方法已经成为目标跟踪领域的主流。以相关滤波为基础的追踪算法，在速度和鲁棒性上实现了很好的效果，但是准确率仍然有较大的提升空间。深度学习类算法能提取出可以较好的区分目标和背景的深度特征，但由于大量的卷积及求逆运算，不能达到实时的要求。基于核相关滤波的KCF目标跟踪算法受到了广泛的关注，利用循环矩阵构造训练样本，使用核函数在傅里叶域内计算岭回归的非线性闭式解，在保证精度的同时，很大程度地提高跟踪速度。深度学习和相关滤波结合，并可以端到端训练的轻型架构的网络DCFNet将Correlation Filter作为独立的一层，使用两个共享参数的CNN组成的孪生网络提取深度特征，在CF层使用循环矩阵构造回归样本，在傅里叶域内计算岭回归的线性闭式解，保留了相关滤波的高效特性，并得到更适合相关滤波区分目标和背景信息的深度特征。本文的工作是解决KCF不能尺度自适应的问题，并将尺度自适应的KCF与DCFNet结合，目的是在保证实时性的同时，进一步提高精度。在ImageNet数据集上训练，在OTB100数据集和部分Temple Color上进行测试，以速度下降16.1%的代价，将成功率提高7.24%。

Object Visual Tracking is a fundamental problem in computer vision. In recent years, there have been many excellent algorithms in which the Correlation Filter and Deep Learning have become the dominant approaches. The object visual tracking algorithm based on Correlation Filter achieves great performance in speed and robustness with the accuracy need to be improved. The tracking methods based on deep learning can extract the deep features which is able to distinguish the target from the background better, but this kind of method can’t operate with real time due to a large number of convolution and inversion operations. The KCF based on Kernel Correlation Filter has received extensive attention from 2014, which constructs the training samples by cyclic shifts operation and the calculates the nonlinear closed-form solution of the ridge regression in the Fourier domain by kernel function. DCFNet constructs a lightweight architecture by combining Deep Learning and Correlation Filter with end-to-end training, in which Correlation Filter is a differentiable layer and a Siamese network of two parameter-shared CNN extracts deep features. Since the derivation is still carried out in Fourier frequency domain, the efficiency property of CF is preserved. The work of this paper is to solve the problem that KCF can't scale adaptively, and combine the scale-adaptive KCF with DCFNet for improving the accuracy while ensuring real-time performance. Trained on the ImageNet dataset and tested on the OTB100 and part of Temple Color test sets, the united tracker increases the accuracy by 7.24% at the cost of a 16.1% speed reduction.