摘要
现有孪生网络目标跟踪器已表现出不错的性能,但对于复杂场景中存在的相似目标误检以及包围框偏移等问题,表现却仍不尽如人意。针对此问题,文中提出了一种新的两阶段孪生目标跟踪算法,该算法由目标提议阶段与任务感知验证阶段组成。第一阶段结合并行空频注意力模块,充分挖掘目标图像的表观特征,增强目标的抗相似物体干扰能力,提高鲁棒性。第二阶段针对检测任务中分类回归任务的差异,对目标分类和位置回归进行任务感知验证,分别获得适用于分类和回归的精准候选框,得到候选目标的识别得分及位置精调。此外,针对训练与测试任务中分类回归计算冲突问题,以及分类回归任务对于正负样本计算存在数量、对象偏差问题,采用GFocal Loss对损失函数进行优化以解决以上问题。实验证明,文中算法在有效性、可靠性以及预期平均重叠率上获得了较大的性能提升,并满足实时跟踪要求。
Siamese-network-based target trackers have shown good performance,but they cannot satisfactorily tackle the problems of false detection on similar targets and bounding box offset in complex scenarios.In this regard,this paper proposes a target-tracking algorithm composed of two stages,the target proposal stage and the task aware verification stage.In the first stage,the parallel space-frequency attention module is integrated to fully mine the apparent features of the target image,and the robustness is improved.In the second stage,since the classification and the regression tasks are different,the task perception verification is carried out separately for target classification and the location regression.Then,the accurate candidate boxes suitable for the classification and those for the regression are obtained,and the identification score and the location fine adjustment of the candidate targets are acquired.In addition,as calculation conflicts exit in classification regression during training and testing,and the quantity and the object errors appear when the positive and the negative samples are calculated,this paper uses the GFocal loss to optimize the loss function.Experiments show that the proposed algorithm can greatly improve effectiveness,reliability and the expected average overlap rate,and is suitable for real-time tracking.
作者
韩光
刘旭辉
刘佶鑫
孙海安
孙宁
HAN Guang;LIU Xuhui;LIU Jixin;SUN Hai'an;SUN Ning(Engineering Research Center of Broadband Wireless Communication Technology,Ministry of Education,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)
出处
《南京邮电大学学报（自然科学版）》
北大核心
2022年第1期62-72,共11页
Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition
基金
国家自然科学基金(61871445,61302156)
江苏省重点研发基金(BE2016001-4)资助项目。
关键词
目标跟踪
孪生网络
并行空频注意力
任务感知
target tracking
siamese network
parallel space-frequency attention
task aware
