基于动态特征注意模型的三分支网络目标跟踪

Triplet Network Based on Dynamic Feature Attention for Object Tracking

针对实际场景中跟踪目标的快速移动、光照变化和尺度变换等问题,提出一种基于动态特征注意模型(DFA)的三分支网络目标跟踪算法,包括:以SiamRPN++跟踪框架为基础,设计具有动态模板分支的在线更新三分支网络,以强化网络提取特征的语义信息,提高模板特征与搜索目标的匹配相似性;设计面向三分支网络训练的样本生成方法,以改变负样本分配方式,提升正、负样本训练的平衡性;设计一种DFA,通过等效自注意和互注意操作增强模板的历史动态特征,实现模板特征的自适应细化,同时利用通道注意力得分控制搜索特征图的权重分配,提高得分图对目标的响应。相对SiamRPN++、SiamBAN等对比算法,所提算法在包含运动模糊、明暗变化和相似背景干扰等场景的OTB100、VOT2018数据集上,获得了最高成功率(71.0%)和最优鲁棒性(0.122),同时可满足实时目标跟踪的要求。

Considering the fast motion,illumination variation,and scale transform of tracking targets in actual scenarios,a triplet network based on a dynamic feature attention(DFA)model for object tracking is proposed to solve these problems.Specifically,on the basis of the SiamRPN++tracking framework,an online update triplet network with dynamic template branches is designed to strengthen the semantic information of extracted features and improve the matching similarity between template features and search features.A sample generation method for the triplet network training is developed to change the allocation of negative samples and improve the balance of positive and negative training samples.Moreover,a DFA model,where the historical dynamic features of the templates are enhanced through equivalent self-attention and mutual attention operation,is designed to achieve the adaptive refinement of template features.Meanwhile,the channel attention score is used to control the weight distribution of the search feature maps,and the response of the score maps is improved.Compared with the state-of-the-art algorithms such as SiamRPN++and SiamBAN,the proposed algorithm has achieved the highest success rate(71.0%)and the best robustness(0.122)on the OTB100 and VOT2018 datasets that contain scenes with motion blur,illumination variation,and similar background interference.This algorithm also can meet the requirement of real-time target tracking.