基于孪生神经网络的跟踪算法是利用离线训练的网络提取目标的特征并进行匹配,从而实现跟踪。在离线训练过程中,网络学到的是相似目标的通用特征,因此当有相似目标干扰时,用这种通用特征表达特定目标将会导致跟踪性能下降,甚至丢失目标。为提高对相似目标的判别能力,通过在线更新网络参数,使网络能够在通用特征的基础上,进一步学到当前目标的特定特征,这样不仅能有效地区分目标与背景,还能消除相似目标的干扰。实验在OTB50和OTB100数据库上进行,结果表明该算法可以提高对网络提取特征的判别力,实现对目标的稳健性跟踪。

Tracking algorithms with Siamese network use the offline training network to extract features from the target for matching and tracking. In the offline training process, the network learns the common features of similar goals. In the case of interference from similar targets, using common features to express specific targets will lead to degradation of tracking performance and even loss of targets. To improve the feature discriminative ability for similar targets, we update the parameters of network online, and make the network further learn the specific characteristics of the current target based on the common features. The proposed method can not only effectively distinguish the target and background, but also eliminate interference from similar targets. We conduct a large number of experiments on the OTB50 and OTB100 databases. The results show that the proposed algorithm can improve the discriminative ability to features extracted by the network and achieve robust tracking of the target.