区域损失函数的孪生网络目标跟踪

针对预训练卷积神经网络提取的深度特征空间分辨率低，快速运动造成运动目标空间细节信息丢失等问题，提出用区域损失函数构建孪生网络的目标跟踪，进一步降低深度特征通道之间的冗余性，并减少高层信息丢失。利用线下预训练的VGG-16卷积神经网络提取深度特征，构成初始深度特征空间。通过区域损失函数构建特征和尺度选择网络，根据反向传播的梯度大小进行特征选择。对筛选后的特征进行拼接，融入到孪生网络中匹配跟踪。在OTB-2013、OTB-2015、VOT2016、TempleColor数据集上与其他算法对比。实验结果表明，该算法在快速运动、低分辨率等场景中表现出较好的跟踪精度和鲁棒性。

Regional loss function based siamese network for object tracking

Due to the low spatial resolution of deep features extracted by pre-trained convolutional neural network, fast motion causes loss of spatial details of a moving object. This paper proposes a method to construct a siamese network for object tracking, so as to reduce the redundancy between the deep feature channels and the loss of high-level information. First, the VGG-16 convolutional neural network is trained offline to extract deep features and form the initial deep feature space. And then, the regional loss function is used to construct the feature and scale selection network. The feature is selected according to the gradient size of back propagation. Further, the selected features are spliced and integrated into the siamese network for matching tracking. By comparing OTB-2013, OTB-2015, VOT2016 and TempleColor benchmark datasets with other algorithms, it shows that the algorithm has preferable precision and robustness in the challenging scenarios such as fast motion and low resolution.