基于显著性区域加权的相关滤波目标跟踪

为了提高跟踪过程中目标位置的定位精度,提出了基于显著性区域加权的相关滤波目标跟踪算法。本文在高效卷积算子跟踪算法(Efficient Convolution Operators for Tracking,ECO)的跟踪框架基础上,首先采用预训练的改进残差网络SE-ResNet来提取不同层的多分辨率特征,充分利用浅层和深层特征的不同特性来增强特征表达,通过因式分解的卷积求出相关滤波的响应图;然后采用背景对像模型来获取目标的显著性图,并使用显著性图来对相关滤波的响应图进行加权,提高定位精度;最后,在视觉目标跟踪(Visual Object Tracking,VOT)竞赛上与8种流行的跟踪算法进行对比,在VOT2016和VOT2017竞赛上的平均重叠期望(Expected Average Overlap,EAO)得分分别达到了0.4157和0.3412,均优于其他算法。实验表明本算法可以有效提升目标跟踪中的定位精度,改善跟踪性能。

Salient region weighted correlation filter for object tracking

To improve the positioning accuracy of target positions in object tracking,an object tracking al_gorithm based on a salient region weighted correlation filter is proposed in this study.Using the tracking framework of efficient convolution operators(ECO)for tracking,we first apply SE-ResNet,which is a pre-trained improved residual network,to extract the multi-resolution features of different layers and fully utilize the different characteristics of the shallow and deep features to enhance feature expression.Next,a background object model is used to obtain a saliency map of the target.The saliency map is then applied to weight the response map of the correlation filter to improve positioning accuracy.Finally,compared with eight popular tracking algorithms employed at the Visual Object Tracking(VOT)challenge,the expected average overlap scores of VOT2016 and VOT2017 are determined to be 0.4157 and 0.3412,respectively,which are better than those of the other algorithms.Experimental results show that the proposed algorithm can effectively improve positioning accuracy and tracking performance.