基于高层语义嵌入的孪生网络跟踪算法

在不加深网络的前提下，为提高孪生网络的特征表达能力，提出基于高层语义嵌入的孪生网络跟踪算法。利用卷积和上采样运算设计了语义嵌入模块，有效融合了深层特征和浅层特征，达到了优化浅层特征的目的，且该模块可以针对任意网络进行灵活的设计与部署。在孪生网络框架下，对AlexNet骨干网络不同层之间添加2个语义嵌入模块。在离线训练阶段进行循环优化，使深层语义信息逐渐转移到较浅的特征层，在跟踪阶段，舍弃语义嵌入模块，仍采用原始的网络结构。实验结果表明：相比于SiamFC，所提算法在OTB2015数据集上精度提高了0.102，成功率提高了0.054。

Siamese network tracking based on high level semantic embedding

In order to improve the feature expression ability of the Siamese network without deepening the network, a Siamese network tracking algorithm was propose based on high-level semantic embedding. First, a semantic embedding module was designed with convolution and up-sampling operations, which effectively integrated deep features with shallow features, thus achieving the purpose of optimizing shallow features, and this module can be flexibly designed and deployed for any network. Then, under the Siamese network framework, two semantic embedding modules were added between different layers of the AlexNet backbone network. Cyclic optimization was carried out in the offline training stage to gradually transfer the deep semantic information to the shallow feature layer. In the tracking stage, the semantic embedding module was abandoned and the original network structure was adopted. The experimental results show that compared with SiamFC on the OTB2015 data set, the accuracy is improved by 0.102 and the success rate is increased by 0.054.