基于区域细化的Siamese网络目标跟踪算法

Siamese network target tracking algorithm based on region refinement

针对快速运动导致跟踪目标尺度变化大、分辨率低等问题,提出了一种基于区域细化的Siamese网络目标跟踪算法。在Siamese网络中引入多尺度特征感知模型,有效提取深层全局通道特征和局部空间特征,准确提取判别性信息;为进一步在搜索区域增强前景,构建区域细化模型,利用经主干网络提取的目标区域特征对搜索区域目标进行甄别,实现由粗到细的跟踪策略,有效增强目标表征能力。将所提算法在OTB100数据集上与现有的一些跟踪算法进行测试。实验结果表明,本文算法在跟踪成功率与跟踪精度方面均取得了良好的表现。同时在低分辨率、形变、光照变化等方面表现出较强的鲁棒性。

In order to solve the problems of large scale variation and low resolution of tracking targets caused by fast movement,a Siamese network target tracking algorithm based on region refinement is proposed.A multi-scale feature sensing model is introduced into Siamese network to effectively extract the deep layer global channel features and local spatial features,so that the algorithm can extract discriminative information accurately.In order to further highlight foreground in the search area,a regional refinement model is constructed,target area features extracted by the backbone network are used to distinguish the search area targets,a coarse to fine tracking strategy is realized,effectively enhance the target representation ability.The proposed algorithm is tested against some existing tracking algorithms on OTB100 datasets.The experimental results show that the proposed algorithm has good performance in tracking success rate and tracking precision.At the same time,it shows strong robustness in low resolution,deformation and illumination variation.