在线目标分类及自适应模板更新的孪生网络跟踪算法

Tracking algorithm of Siamese network based on online target classification and adaptive template update

针对孪生网络跟踪算法在离线训练阶段学习被跟踪目标和其他对象的嵌入式特征,而这些特征缺少特定于目标的上下文信息,使跟踪算法的稳健性较差的问题,以SiamRPN++作为基准算法,提出了在线目标分类及自适应模板更新的孪生网络跟踪算法。首先,在离线训练阶段设计了互相关特征图监督模块,以学习更具判别力的嵌入式特征;其次,在线跟踪阶段设计了包含注意力机制的在线目标分类模块,在该模块中使用在线滤波器更新策略滤除背景噪声干扰;最后,设计了一种自适应模板更新模块,使用UpdateNet更新目标模板信息。在VOT2018、VOT2019这2个标准数据集上的实验结果验证了所提算法的有效性,相比基准算法SiamRPN++分别带来13.5%和18.2%(EAO)的性能提升。

Aiming at the problem that tracking algorithm of Siamese network learned the embedded features of the tracked target and the object in the offline training stage,and these embedded features often lacked the target-specific context information,which made these tracking algorithms less robust,a tracking algorithm of the Siamese network based on online target classification and adaptive template update was proposed,which used SiamRPN++as the baseline algorithm.Firstly,a cross-correlation feature map supervision module for classification was designed in the offline train-ing phase to learn more discriminative embedded features.Secondly,an online target classification module that included an attention mechanism in the online tracking phase was designed,and the online update filter strategy in the module was used to filter out the background noise.Finally,an adaptive template update module was designed to update the target template information using the UpdateNet.The results of experiments on VOT2018 and VOT2019 datasets verify the ef-fectiveness of the proposed algorithm,which brings 13.5%and 18.2%(EAO)improvement respectively compared with the baseline algorithm SiamRPN++.