融合卷积神经网络的核相关滤波视觉目标跟随算法研究

Visual Object Tracking Combining Kernel Correlation Filter andConvolutional Neural Network

近几年,目标跟随技术逐渐成为研究的热点;核相关滤波跟踪算法通过循环矩阵构造训练样本,将时域的卷积转换到频域的点乘完成滤波器的训练,降低计算复杂度,跟踪速度较快;卷积神经网络模型深度特征表征能力较强,可以充分利用图像信息,跟踪精度较高;将两种算法优势互补,构造一种卷积神经网络与核相关滤波算法融合型改进算法;即在线下阶段训练模型,分层提取孪生网络的深度特征,然后通过相关滤波器快速计算出最大响应图,预测目标所在位置;因此,改进后的算法在保持核相关滤波跟踪算法实时性的同时,可以大幅提高跟踪精度。

In recent years,object tracking technology has gradually become a research hotspot.The deep feature representation ability of the convolutional neural network model is strong,which can make full use of image information,and can distinguish categories.Meanwhile,the kernel correlation filter tracking algorithm converts the convolution in the time domain to the dot product in the frequency domain,which reduces the calculation complexity and significantly improves the calculation speed.Complementing the advantages of the two algorithms,a new tracking algorithm that combines convolutional neural network and kernel correlation filter algorithm is proposed.Specifically,the deep features of the Siamese network are extracted hierarchically to construct training the model offline.Then kernel correlation filter algorithm quickly calculates the target and predicts the location of the target.Therefore,the proposed algorithm can improve the performance of tracking accuracy.