【摘要】 目标尺度变化和低分辨率的复杂场景往往会影响目标跟踪算法的性能进而导致跟踪精度下降。针对此问题，提出了一种基于深度像素级特征的孪生网络目标跟踪方法。引入像素级特征融合方法对目标模板和搜索区域的多层特征进行融合、设计基于残差网络和拓扑结构的特征深层提取模块、依据判据筛选历史信息得到合适模板特征进行模板更新。实验结果表明，所提改进算法在VOT2018数据集上比基础算法的EAO值提升了5.31%,准确率提升了0.83%,鲁棒性提升了3.85%;在OTB100数据集上，所提算法精确率为91.4%,成功率为71.7%,与基础算法相比，精确率提升了3.28%,成功率提升了5.13%。更多还原

【Abstract】 Complex scenes, such as target scale changes and low resolution often affect the performance of tracking algorithm, which leads to the decline of tracking accuracy. To solve this problem, a Siamese Network tracking method based on deep pixel wise feature is proposed. The pixel wise feature fusion is introduced to fuse the multi-layer features of the template and the search region. The deep feature extraction module based on residual network and topological structure is designed, and the historical information is filtered to update the template. The experimental results show that on VOT2018,the EAO value of the improvement algorithm is improved by 5.31%,the accuracy is improved by 0.83%,and the robustness is improved by 3.85%. On OTB100,the accuracy of this algorithm is 91.4% and the success rate is 71.7%. Compared with the basic algorithm, the accuracy is improved by 3.28% and the success rate is improved by 5.13%.更多还原