基于非对称卷积的孪生网络视觉跟踪算法

Siamese Network Visual Tracking Based on Asymmetric Convolution

针对孪生网络对旋转变化目标特征表达能力不足的问题,该文提出了基于非对称卷积的孪生网络跟踪算法。首先利用卷积核的可加性构建非对称卷积核组,可以将其应用于任意卷积核大小的已有网络结构。接着在孪生网络跟踪框架下,对AlexNet的卷积模块进行替换,并在训练和跟踪阶段对网络进行分别设计。最后在网络的末端并联地添加3个非对称卷积核,分别经过相关运算后得到3个响应图,进行加权融合后选取最大值即为目标的位置。实验结果表明,相比于SiamFC,在OTB2015数据集上精度提高了8.7%,成功率提高了4.5%。

In order to solve the problem that the Siamese network can not express the rotating target,a Siamese network tracking algorithm based on asymmetric convolution is proposed.Firstly,asymmetric convolution kernels are constructed,which can be applied to existing networks.Then,under the framework of Siamese network,the convolution module of AlexNet is replaced,and the network is designed separately in the training and tracking stages.Finally,three asymmetric convolution kernels are added in parallel in the last layer of the network,and the maximum value is selected as the target position after the three response maps are weighted fused.The experimental results show that compared with SiamFC,the accuracy and success rate are improved by 8.7%and 4.5%on OTB2015 dataset,respectively.