Abstract:

There are problems in object tracking based on siamese networks, such as insufficient feature information, tracking efficiency needing to be improved, and long training time on large datasets. In order to solve the above problems, efficient object tracking based on feature fusion and training acceleration is proposed. The reference feature level of backbone network is increased, down-sampling is reduced, and multi-level reference feature maps are fused with the purpose of extracting deeper and more abundant semantic targets information. Depth-wise cross correlation is used to get RoWs （Response of Candidate Windows）, in which the Region Proposal Network（RPN） is constructed. By balancing the number of positive and negative anchors, the performance of siamese networks is more efficient and steady. When training siamese networks on large datasets, uniformly sliding shift sampling is used instead of randomly shift sampling, which significantly accelerates training speed at the time of suppressing center bias phenomenon. Experimental evaluation results on tracking benchmark VOT2018 show that the proposed algorithm has the best tracking performance compared with all reference mainstream object tracking algorithms.

Key words: object tracking, siamese network, reference feature map fusion, depth-wise cross correlation, Response of Candidate Windows（RoWs）, balancing anchors, shift sampling, training acceleration

摘要：

基于孪生网络的目标跟踪，存在特征信息欠丰富，跟踪效率有待提高，大型数据集上训练时间长等问题。针对上述问题，提出特征融合与训练加速的高效目标跟踪。增加主干网络参考特征层级，减小下采样，融合多层级参考特征图，提取目标更深度、丰富的语义信息。深度互相关操作得到候选窗口响应（Response of Candidate Windows，RoWs），在其中构建区域建议网络（Region Proposal Network，RPN），通过权衡正负锚点的数量比，使孪生网络性能更加高效、稳定。大型数据集训练孪生网络时，使用均匀滑动漂移采样，代替随机漂移采样算法，在抑制中心偏置现象的同时，显著加快了孪生网络的训练速度。跟踪基准VOT2018上的评估实验结果表明，与所有参考的主流目标跟踪算法相比，所提算法具有最佳的跟踪性能。

关键词: 目标跟踪, 孪生网络, 参考特征图融合, 深度互相关, 候选窗口响应（RoWs）, 权衡锚点, 漂移采样, 训练加速