基于多分支和重参数化的孪生网络跟踪算法

Siamese Network Tracking Algorithm Based on Diverse Branch Block and Reparameterization

针对孪生网络对尺度变化目标特征表达能力不足的问题,本文使用不同尺寸的卷积、池化分支和剪枝操作构成多分支结构,以提高特征的鲁棒性并保证孪生网络的平移不变性。针对多分支结构带来参数量过多的问题,本文在跟踪阶段将多分支结构重参数化为单一的卷积,有效减少跟踪阶段时间成本。实验结果表明:本文提出的算法相比于SiamFC,在OTB100数据集上,其精度、成功率和跟踪速度分别提高了5.1%、3%、30FPS,在UAV123和Temple-Color-128数据集上跟踪精度和成功率均有所提高。

Aiming at the problem that the Siamese network has insufficient ability to express the features of scale-varying targets,a multi-branch structure is constructed by using convolution,pooling branches and pruning operations of different sizes to improve the robustness of features and ensure the translation invariance of the Siamese network.Aiming at the problem that the multi-branch structure brings too many parameters,the multi-branch structure is reparameterized into a single convolution in the tracking stage,which effectively reduces the time cost in the tracking stage.The experimental results show that compared with SiamFC,the accuracy,success rate and tracking speed of the proposed algorithm on the OTB100 datasets are improved by 5.1%,3%and 30 FPS,respectively.The tracking accuracy and success rate are improved on the UAV123 and Temple-Color-128 datasets.